fix: address review feedback — race condition, GPU safety, task GC

- Add threading lock to get_tts_backend_for_engine() to prevent race
  condition where concurrent requests could create duplicate backend
  instances (double-checked locking pattern)
- Fix LuxTTS generate: call .detach().cpu() before .numpy() so it
  works on GPU/MPS devices, not just CPU
- Store background download tasks in a module-level set to prevent
  garbage collection before completion (asyncio.create_task fire-and-
  forget pattern)
- Deduplicate cache_key computation in LuxTTS create_voice_prompt
- Prefix unused sr variable with underscore

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.1.12
+current_version = 0.1.13
 commit = True
 tag = True
 tag_name = v{new_version}

.github/workflows/build-cuda.yml

@@ -0,0 +1,73 @@
+name: Build CUDA Backend
+
+on:
+  workflow_dispatch:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  build-cuda-windows:
+    runs-on: windows-latest
+    permissions:
+      contents: write
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: "pip"
+
+      - name: Install Python dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install pyinstaller
+          pip install -r backend/requirements.txt
+
+      - name: Install PyTorch with CUDA 12.1
+        run: |
+          pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-deps
+          pip install torchaudio --index-url https://download.pytorch.org/whl/cu121
+
+      - name: Verify CUDA support in torch
+        run: |
+          python -c "import torch; print(f'CUDA available in build: {torch.cuda.is_available()}'); print(f'CUDA version: {torch.version.cuda}')"
+
+      - name: Build CUDA server binary
+        shell: bash
+        working-directory: backend
+        run: python build_binary.py --cuda
+
+      - name: Split binary for GitHub Releases
+        shell: bash
+        run: |
+          python scripts/split_binary.py \
+            backend/dist/voicebox-server-cuda.exe \
+            --output release-assets/
+
+      - name: Upload split parts to GitHub Release
+        if: startsWith(github.ref, 'refs/tags/')
+        uses: softprops/action-gh-release@v1
+        with:
+          files: |
+            release-assets/voicebox-server-cuda.part*.exe
+            release-assets/voicebox-server-cuda.sha256
+            release-assets/voicebox-server-cuda.manifest
+          draft: true
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Upload binary as workflow artifact (for testing)
+        uses: actions/upload-artifact@v4
+        with:
+          name: voicebox-server-cuda-windows
+          path: backend/dist/voicebox-server-cuda.exe
+          retention-days: 7
+
+  # Linux CUDA build can be added later with:
+  # build-cuda-linux:
+  #   runs-on: ubuntu-22.04
+  #   ...

.github/workflows/build-windows.yml

@@ -0,0 +1,63 @@
+name: Build Windows
+
+on:
+  workflow_dispatch:
+
+jobs:
+  build-windows:
+    permissions:
+      contents: write
+    runs-on: windows-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: "pip"
+
+      - name: Install Python dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install pyinstaller
+          pip install -r backend/requirements.txt
+
+      - name: Build Python server
+        shell: bash
+        run: |
+          cd backend
+          python build_binary.py
+
+          PLATFORM=$(rustc --print host-tuple)
+          mkdir -p ../tauri/src-tauri/binaries
+          cp dist/voicebox-server.exe ../tauri/src-tauri/binaries/voicebox-server-${PLATFORM}.exe
+          echo "Built voicebox-server-${PLATFORM}.exe"
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+
+      - name: Install Rust stable
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Rust cache
+        uses: swatinem/rust-cache@v2
+        with:
+          workspaces: "./tauri/src-tauri -> target"
+
+      - name: Install dependencies
+        run: bun install
+
+      - uses: tauri-apps/tauri-action@v0
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          projectPath: tauri
+          tagName: v__VERSION__
+          releaseName: "voicebox v__VERSION__ (test build)"
+          releaseBody: "Test build for audio export fix"
+          releaseDraft: true
+          prerelease: true
+          args: ""
+          includeUpdaterJson: false

CONTRIBUTING.md

@@ -32,7 +32,26 @@ Thank you for your interest in contributing to Voicebox! This document provides
 
 ### Development Setup
 
-**Using the Makefile (recommended for macOS/Linux):** Run `make setup` to install all dependencies, then `make dev` to start development servers. See `make help` for all available commands.
+**Using `just` (recommended):**
+
+Install [just](https://github.com/casey/just) (`brew install just` or `cargo install just`), then:
+
+```bash
+just setup   # creates venv, installs Python + JS deps
+just dev     # starts backend + desktop app in one terminal
+```
+
+Other useful commands:
+
+```bash
+just dev-web       # backend + web app (no Tauri/Rust build)
+just dev-backend   # backend only
+just kill          # stop all dev processes
+just clean-all     # nuke everything and start fresh
+just --list        # see all available commands
+```
+
+**Using the Makefile:** Run `make setup` then `make dev`. See `make help` for all commands.
 
 **Manual setup (required for Windows):**
 

README.md

@@ -80,10 +80,10 @@ Voicebox is available now for macOS and Windows.
 
 | Platform | Download |
 |----------|----------|
-| macOS (Apple Silicon) | [voicebox_aarch64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_aarch64.app.tar.gz) |
-| macOS (Intel) | [voicebox_x64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_x64.app.tar.gz) |
-| Windows (MSI) | [voicebox_0.1.0_x64_en-US.msi](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_0.1.0_x64_en-US.msi) |
-| Windows (Setup) | [voicebox_0.1.0_x64-setup.exe](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_0.1.0_x64-setup.exe) |
+| macOS (Apple Silicon) | [Voicebox_aarch64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/latest/download/Voicebox_aarch64.app.tar.gz) |
+| macOS (Intel) | [Voicebox_x64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/latest/download/Voicebox_x64.app.tar.gz) |
+| Windows (MSI) | [Latest Windows MSI](https://github.com/jamiepine/voicebox/releases/latest) |
+| Windows (Setup) | [Latest Windows Setup](https://github.com/jamiepine/voicebox/releases/latest) |
 
 > **Linux builds coming soon** — Currently blocked by GitHub runner disk space limitations.
 
@@ -225,42 +225,21 @@ Voicebox aims to be the **one-stop shop for everything voice** — cloning, synt
 
 See [CONTRIBUTING.md](CONTRIBUTING.md) for detailed setup and contribution guidelines.
 
-**Using the Makefile (recommended):** Run `make help` to see all available commands for setup, development, building, and testing.
-
 ### Quick Start
 
-**With Makefile (Unix/macOS/Linux):**
-
 ```bash
-# Clone the repo
-git clone https://github.com/voicebox-sh/voicebox.git
+git clone https://github.com/jamiepine/voicebox.git
 cd voicebox
 
-# Setup everything
-make setup
-
-# Start development
-make dev
+just setup   # creates Python venv, installs all deps
+just dev     # starts backend + desktop app
 ```
 
-**Manual setup (all platforms):**
+Install [just](https://github.com/casey/just): `brew install just` or `cargo install just`. Run `just --list` to see all commands.
 
-```bash
-# Clone the repo
-git clone https://github.com/voicebox-sh/voicebox.git
-cd voicebox
+Also available via Makefile: `make setup && make dev` (run `make help` for all commands).
 
-# Install dependencies
-bun install
-
-# Install Python dependencies
-cd backend && pip install -r requirements.txt && cd ..
-
-# Start development
-bun run dev
-```
-
-**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org). 
+**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org), [XCode on macOS](https://developer.apple.com/xcode/).
 
 **Performance:** 
 - **Apple Silicon (M1/M2/M3)**: Uses MLX backend with native Metal acceleration for 4-5x faster inference

app/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/app",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "private": true,
   "type": "module",
   "scripts": {

app/src/components/AudioTab/AudioTab.tsx

@@ -124,6 +124,13 @@ export function AudioTab() {
     );
   }
 
+  const handleChannelDelete = async (e, channelId) => {
+    e.stopPropagation();
+    if (await confirm('Delete this channel?')) {
+      deleteChannel.mutate(channelId);
+    }
+  }
+
   const allChannels = channels || [];
   const allDevices = devices || [];
   const selectedChannel = selectedChannelId
@@ -241,12 +248,7 @@ export function AudioTab() {
                             variant="ghost"
                             size="sm"
                             className="h-8 w-8 p-0"
-                            onClick={(e) => {
-                              e.stopPropagation();
-                              if (confirm('Delete this channel?')) {
-                                deleteChannel.mutate(channel.id);
-                              }
-                            }}
+                            onClick={(e) => handleChannelDelete(e, channel.id)}
                           >
                             <Trash2 className="h-4 w-4" />
                           </Button>

app/src/components/Generation/FloatingGenerateBox.tsx

@@ -316,7 +316,7 @@ export function FloatingGenerateBox({
                   </span>
                 </div>
                 <AnimatePresence>
-                  {isExpanded && (
+                  {isExpanded && form.watch('engine') !== 'luxtts' && (
                     <motion.div
                       initial={{ opacity: 0, scale: 0.8 }}
                       animate={{ opacity: 1, scale: 1 }}
@@ -402,30 +402,41 @@ export function FloatingGenerateBox({
                     )}
                   />
 
-                  <FormField
-                    control={form.control}
-                    name="modelSize"
-                    render={({ field }) => (
-                      <FormItem className="flex-1 space-y-0">
-                        <Select onValueChange={field.onChange} defaultValue={field.value}>
-                          <FormControl>
-                            <SelectTrigger className="h-8 text-xs bg-card border-border rounded-full hover:bg-background/50 transition-all">
-                              <SelectValue />
-                            </SelectTrigger>
-                          </FormControl>
-                          <SelectContent>
-                            <SelectItem value="1.7B" className="text-xs text-muted-foreground">
-                              Qwen3-TTS 1.7B
-                            </SelectItem>
-                            <SelectItem value="0.6B" className="text-xs text-muted-foreground">
-                              Qwen3-TTS 0.6B
-                            </SelectItem>
-                          </SelectContent>
-                        </Select>
-                        <FormMessage className="text-xs" />
-                      </FormItem>
-                    )}
-                  />
+                  <FormItem className="flex-1 space-y-0">
+                    <Select
+                      value={
+                        form.watch('engine') === 'luxtts'
+                          ? 'luxtts'
+                          : `qwen:${form.watch('modelSize') || '1.7B'}`
+                      }
+                      onValueChange={(value) => {
+                        if (value === 'luxtts') {
+                          form.setValue('engine', 'luxtts');
+                        } else {
+                          const [, modelSize] = value.split(':');
+                          form.setValue('engine', 'qwen');
+                          form.setValue('modelSize', modelSize as '1.7B' | '0.6B');
+                        }
+                      }}
+                    >
+                      <FormControl>
+                        <SelectTrigger className="h-8 text-xs bg-card border-border rounded-full hover:bg-background/50 transition-all">
+                          <SelectValue />
+                        </SelectTrigger>
+                      </FormControl>
+                      <SelectContent>
+                        <SelectItem value="qwen:1.7B" className="text-xs text-muted-foreground">
+                          Qwen3-TTS 1.7B
+                        </SelectItem>
+                        <SelectItem value="qwen:0.6B" className="text-xs text-muted-foreground">
+                          Qwen3-TTS 0.6B
+                        </SelectItem>
+                        <SelectItem value="luxtts" className="text-xs text-muted-foreground">
+                          LuxTTS
+                        </SelectItem>
+                      </SelectContent>
+                    </Select>
+                  </FormItem>
                 </div>
               </motion.div>
             </AnimatePresence>

app/src/components/Generation/GenerationForm.tsx

@@ -76,29 +76,67 @@ export function GenerationForm() {
               )}
             />
 
-            <FormField
-              control={form.control}
-              name="instruct"
-              render={({ field }) => (
-                <FormItem>
-                  <FormLabel>Delivery Instructions (optional)</FormLabel>
-                  <FormControl>
-                    <Textarea
-                      placeholder="e.g. Speak slowly with emphasis, Warm and friendly tone, Professional and authoritative..."
-                      className="min-h-[80px]"
-                      {...field}
-                    />
-                  </FormControl>
-                  <FormDescription>
-                    Natural language instructions to control speech delivery (tone, emotion, pace).
-                    Max 500 characters
-                  </FormDescription>
-                  <FormMessage />
-                </FormItem>
-              )}
-            />
+            {form.watch('engine') !== 'luxtts' && (
+              <FormField
+                control={form.control}
+                name="instruct"
+                render={({ field }) => (
+                  <FormItem>
+                    <FormLabel>Delivery Instructions (optional)</FormLabel>
+                    <FormControl>
+                      <Textarea
+                        placeholder="e.g. Speak slowly with emphasis, Warm and friendly tone, Professional and authoritative..."
+                        className="min-h-[80px]"
+                        {...field}
+                      />
+                    </FormControl>
+                    <FormDescription>
+                      Natural language instructions to control speech delivery (tone, emotion,
+                      pace). Max 500 characters
+                    </FormDescription>
+                    <FormMessage />
+                  </FormItem>
+                )}
+              />
+            )}
 
             <div className="grid gap-4 md:grid-cols-3">
+              <FormItem>
+                <FormLabel>Model</FormLabel>
+                <Select
+                  value={
+                    form.watch('engine') === 'luxtts'
+                      ? 'luxtts'
+                      : `qwen:${form.watch('modelSize') || '1.7B'}`
+                  }
+                  onValueChange={(value) => {
+                    if (value === 'luxtts') {
+                      form.setValue('engine', 'luxtts');
+                    } else {
+                      const [, modelSize] = value.split(':');
+                      form.setValue('engine', 'qwen');
+                      form.setValue('modelSize', modelSize as '1.7B' | '0.6B');
+                    }
+                  }}
+                >
+                  <FormControl>
+                    <SelectTrigger>
+                      <SelectValue />
+                    </SelectTrigger>
+                  </FormControl>
+                  <SelectContent>
+                    <SelectItem value="qwen:1.7B">Qwen3-TTS 1.7B</SelectItem>
+                    <SelectItem value="qwen:0.6B">Qwen3-TTS 0.6B</SelectItem>
+                    <SelectItem value="luxtts">LuxTTS</SelectItem>
+                  </SelectContent>
+                </Select>
+                <FormDescription>
+                  {form.watch('engine') === 'luxtts'
+                    ? 'Fast, English-focused'
+                    : 'Multi-language, two sizes'}
+                </FormDescription>
+              </FormItem>
+
               <FormField
                 control={form.control}
                 name="language"
@@ -124,29 +162,6 @@ export function GenerationForm() {
                 )}
               />
 
-              <FormField
-                control={form.control}
-                name="modelSize"
-                render={({ field }) => (
-                  <FormItem>
-                    <FormLabel>Model Size</FormLabel>
-                    <Select onValueChange={field.onChange} defaultValue={field.value}>
-                      <FormControl>
-                        <SelectTrigger>
-                          <SelectValue />
-                        </SelectTrigger>
-                      </FormControl>
-                      <SelectContent>
-                        <SelectItem value="1.7B">Qwen TTS 1.7B (Higher Quality)</SelectItem>
-                        <SelectItem value="0.6B">Qwen TTS 0.6B (Faster)</SelectItem>
-                      </SelectContent>
-                    </Select>
-                    <FormDescription>Larger models produce better quality</FormDescription>
-                    <FormMessage />
-                  </FormItem>
-                )}
-              />
-
               <FormField
                 control={form.control}
                 name="seed"
@@ -170,11 +185,7 @@ export function GenerationForm() {
               />
             </div>
 
-            <Button
-              type="submit"
-              className="w-full"
-              disabled={isPending || !selectedProfileId}
-            >
+            <Button type="submit" className="w-full" disabled={isPending || !selectedProfileId}>
               {isPending ? (
                 <>
                   <Loader2 className="mr-2 h-4 w-4 animate-spin" />

app/src/components/ServerSettings/GpuAcceleration.tsx

@@ -0,0 +1,387 @@
+import { useQuery, useQueryClient } from '@tanstack/react-query';
+import { AlertCircle, Cpu, Download, Loader2, RotateCw, Trash2, Zap } from 'lucide-react';
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { Badge } from '@/components/ui/badge';
+import { Button } from '@/components/ui/button';
+import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card';
+import { Progress } from '@/components/ui/progress';
+import { apiClient } from '@/lib/api/client';
+import type { CudaDownloadProgress } from '@/lib/api/types';
+import { useServerHealth } from '@/lib/hooks/useServer';
+import { usePlatform } from '@/platform/PlatformContext';
+import { useServerStore } from '@/stores/serverStore';
+
+type RestartPhase = 'idle' | 'stopping' | 'waiting' | 'ready';
+
+export function GpuAcceleration() {
+  const platform = usePlatform();
+  const queryClient = useQueryClient();
+  const serverUrl = useServerStore((state) => state.serverUrl);
+  const { data: health } = useServerHealth();
+
+  const [restartPhase, setRestartPhase] = useState<RestartPhase>('idle');
+  const [error, setError] = useState<string | null>(null);
+  const [downloadProgress, setDownloadProgress] = useState<CudaDownloadProgress | null>(null);
+  const healthPollRef = useRef<ReturnType<typeof setInterval> | null>(null);
+
+  // Query CUDA backend status
+  const {
+    data: cudaStatus,
+    isLoading: cudaStatusLoading,
+    refetch: refetchCudaStatus,
+  } = useQuery({
+    queryKey: ['cuda-status', serverUrl],
+    queryFn: () => apiClient.getCudaStatus(),
+    refetchInterval: cudaStatusLoading ? false : 10000,
+    retry: 1,
+    enabled: !!health, // Only fetch when backend is reachable
+  });
+
+  // Derived state
+  const isCurrentlyCuda = health?.backend_variant === 'cuda';
+  const cudaAvailable = cudaStatus?.available ?? false;
+  const cudaDownloading = cudaStatus?.downloading ?? false;
+
+  // Clean up health poll on unmount
+  useEffect(() => {
+    return () => {
+      if (healthPollRef.current) {
+        clearInterval(healthPollRef.current);
+        healthPollRef.current = null;
+      }
+    };
+  }, []);
+
+  // SSE progress tracking during download
+  useEffect(() => {
+    if (!cudaDownloading || !serverUrl) {
+      return;
+    }
+
+    const eventSource = new EventSource(`${serverUrl}/backend/cuda-progress`);
+
+    eventSource.onmessage = (event) => {
+      try {
+        const data = JSON.parse(event.data) as CudaDownloadProgress;
+        setDownloadProgress(data);
+
+        if (data.status === 'complete') {
+          eventSource.close();
+          setDownloadProgress(null);
+          refetchCudaStatus();
+        } else if (data.status === 'error') {
+          eventSource.close();
+          setError(data.error || 'Download failed');
+          setDownloadProgress(null);
+          refetchCudaStatus();
+        }
+      } catch (e) {
+        console.error('Error parsing CUDA progress event:', e);
+      }
+    };
+
+    eventSource.onerror = () => {
+      eventSource.close();
+    };
+
+    return () => {
+      eventSource.close();
+    };
+  }, [cudaDownloading, serverUrl, refetchCudaStatus]);
+
+  // Start aggressive health polling during restart
+  const startHealthPolling = useCallback(() => {
+    if (healthPollRef.current) return;
+
+    healthPollRef.current = setInterval(async () => {
+      try {
+        const result = await apiClient.getHealth();
+        if (result.status === 'healthy') {
+          // Server is back up
+          if (healthPollRef.current) {
+            clearInterval(healthPollRef.current);
+            healthPollRef.current = null;
+          }
+          setRestartPhase('ready');
+          // Invalidate all queries to refresh UI
+          queryClient.invalidateQueries();
+          // Reset after a moment
+          setTimeout(() => setRestartPhase('idle'), 2000);
+        }
+      } catch {
+        // Server still down, keep polling
+      }
+    }, 1000);
+  }, [queryClient]);
+
+  const handleDownload = async () => {
+    setError(null);
+    try {
+      await apiClient.downloadCudaBackend();
+      refetchCudaStatus();
+    } catch (e: unknown) {
+      const msg = e instanceof Error ? e.message : 'Failed to start download';
+      if (msg.includes('already downloaded')) {
+        refetchCudaStatus();
+      } else {
+        setError(msg);
+      }
+    }
+  };
+
+  const handleRestart = async () => {
+    setError(null);
+    setRestartPhase('stopping');
+
+    try {
+      setRestartPhase('waiting');
+      startHealthPolling();
+      await platform.lifecycle.restartServer();
+      // Invoke resolved — server is likely ready. Stop polling and refresh.
+      if (healthPollRef.current) {
+        clearInterval(healthPollRef.current);
+        healthPollRef.current = null;
+      }
+      setRestartPhase('ready');
+      queryClient.invalidateQueries();
+      setTimeout(() => setRestartPhase('idle'), 2000);
+    } catch (e: unknown) {
+      setRestartPhase('idle');
+      if (healthPollRef.current) {
+        clearInterval(healthPollRef.current);
+        healthPollRef.current = null;
+      }
+      setError(e instanceof Error ? e.message : 'Restart failed');
+    }
+  };
+
+  const handleSwitchToCpu = async () => {
+    // To switch to CPU: delete the CUDA binary, then restart.
+    // start_server always prefers CUDA if present, so we must remove it first.
+    setError(null);
+    setRestartPhase('stopping');
+
+    try {
+      await apiClient.deleteCudaBackend();
+      setRestartPhase('waiting');
+      startHealthPolling();
+      await platform.lifecycle.restartServer();
+      // Invoke resolved — server is likely ready
+      if (healthPollRef.current) {
+        clearInterval(healthPollRef.current);
+        healthPollRef.current = null;
+      }
+      setRestartPhase('ready');
+      queryClient.invalidateQueries();
+      setTimeout(() => setRestartPhase('idle'), 2000);
+    } catch (e: unknown) {
+      setRestartPhase('idle');
+      if (healthPollRef.current) {
+        clearInterval(healthPollRef.current);
+        healthPollRef.current = null;
+      }
+      setError(e instanceof Error ? e.message : 'Failed to switch to CPU');
+      refetchCudaStatus();
+    }
+  };
+
+  const handleDelete = async () => {
+    setError(null);
+    try {
+      await apiClient.deleteCudaBackend();
+      refetchCudaStatus();
+    } catch (e: unknown) {
+      setError(e instanceof Error ? e.message : 'Failed to delete CUDA backend');
+    }
+  };
+
+  const formatBytes = (bytes: number): string => {
+    if (bytes === 0) return '0 B';
+    const k = 1024;
+    const sizes = ['B', 'KB', 'MB', 'GB'];
+    const i = Math.floor(Math.log(bytes) / Math.log(k));
+    return `${(bytes / k ** i).toFixed(1)} ${sizes[i]}`;
+  };
+
+  // Don't render until health data is available
+  if (!health) return null;
+
+  // If the system already has native GPU (MPS, etc.), only show info - no CUDA needed
+  const hasNativeGpu =
+    health.gpu_available &&
+    !isCurrentlyCuda &&
+    health.gpu_type &&
+    !health.gpu_type.includes('CUDA');
+
+  return (
+    <Card>
+      <CardHeader>
+        <CardTitle className="flex items-center gap-2">
+          <Zap className="h-4 w-4" />
+          GPU Acceleration
+        </CardTitle>
+      </CardHeader>
+      <CardContent className="space-y-4">
+        {/* Current status */}
+        <div className="flex items-center justify-between">
+          <div className="space-y-1">
+            <div className="text-sm font-medium">Backend</div>
+            <div className="text-sm text-muted-foreground">
+              {isCurrentlyCuda ? 'CUDA (GPU accelerated)' : 'CPU'}
+            </div>
+          </div>
+          <Badge variant={isCurrentlyCuda ? 'default' : 'secondary'}>
+            {isCurrentlyCuda ? (
+              <>
+                <Zap className="h-3 w-3 mr-1" /> CUDA
+              </>
+            ) : (
+              <>
+                <Cpu className="h-3 w-3 mr-1" /> CPU
+              </>
+            )}
+          </Badge>
+        </div>
+
+        {/* GPU info from health */}
+        {health.gpu_type && (
+          <div className="space-y-1">
+            <div className="text-sm font-medium">GPU</div>
+            <div className="text-sm text-muted-foreground">{health.gpu_type}</div>
+            {health.vram_used_mb != null && (
+              <div className="text-xs text-muted-foreground">
+                VRAM: {health.vram_used_mb.toFixed(0)} MB used
+              </div>
+            )}
+          </div>
+        )}
+
+        {/* Native GPU detected - no CUDA download needed */}
+        {hasNativeGpu && (
+          <div className="p-3 rounded-lg bg-accent/10 border border-accent/20">
+            <div className="text-sm">
+              Your system uses <strong>{health.gpu_type}</strong> for acceleration. No additional
+              downloads needed.
+            </div>
+          </div>
+        )}
+
+        {/* CUDA download section - only show when native GPU is NOT detected (i.e., Windows/Linux NVIDIA users) */}
+        {!hasNativeGpu && (
+          <>
+            {/* Download progress */}
+            {cudaDownloading && downloadProgress && (
+              <div className="space-y-2">
+                <div className="flex items-center justify-between text-sm">
+                  <div className="flex items-center gap-2">
+                    <Loader2 className="h-4 w-4 animate-spin" />
+                    <span>{downloadProgress.filename || 'Downloading CUDA backend...'}</span>
+                  </div>
+                  {downloadProgress.total > 0 && (
+                    <span className="text-muted-foreground">
+                      {downloadProgress.progress.toFixed(1)}%
+                    </span>
+                  )}
+                </div>
+                {downloadProgress.total > 0 && (
+                  <>
+                    <Progress value={downloadProgress.progress} className="h-2" />
+                    <div className="text-xs text-muted-foreground">
+                      {formatBytes(downloadProgress.current)} /{' '}
+                      {formatBytes(downloadProgress.total)}
+                    </div>
+                  </>
+                )}
+              </div>
+            )}
+
+            {/* Restart in progress */}
+            {restartPhase !== 'idle' && (
+              <div className="flex items-center gap-2 p-3 rounded-lg bg-primary/5 border">
+                <Loader2 className="h-4 w-4 animate-spin" />
+                <span className="text-sm">
+                  {restartPhase === 'stopping' && 'Stopping server...'}
+                  {restartPhase === 'waiting' && 'Restarting server...'}
+                  {restartPhase === 'ready' && 'Server restarted successfully!'}
+                </span>
+              </div>
+            )}
+
+            {/* Error display */}
+            {error && (
+              <div className="flex items-center gap-2 text-sm text-destructive">
+                <AlertCircle className="h-4 w-4 shrink-0" />
+                <span>{error}</span>
+              </div>
+            )}
+
+            {/* Actions */}
+            {restartPhase === 'idle' && !cudaDownloading && (
+              <div className="space-y-2">
+                {/* Not downloaded yet - show download button */}
+                {!cudaAvailable && (
+                  <div className="space-y-3">
+                    <p className="text-sm text-muted-foreground">
+                      Download the CUDA backend (~2.4 GB) for NVIDIA GPU acceleration. Requires an
+                      NVIDIA GPU with CUDA support.
+                    </p>
+                    <Button onClick={handleDownload} className="w-full" size="sm">
+                      <Download className="h-4 w-4 mr-2" />
+                      Download CUDA Backend
+                    </Button>
+                  </div>
+                )}
+
+                {/* Downloaded but not active - show switch button */}
+                {cudaAvailable && !isCurrentlyCuda && platform.metadata.isTauri && (
+                  <div className="space-y-3">
+                    <p className="text-sm text-muted-foreground">
+                      CUDA backend is downloaded and ready. Restart the server to enable GPU
+                      acceleration.
+                    </p>
+                    <Button onClick={handleRestart} className="w-full" size="sm">
+                      <RotateCw className="h-4 w-4 mr-2" />
+                      Switch to CUDA Backend
+                    </Button>
+                  </div>
+                )}
+
+                {/* Currently active - show switch back to CPU */}
+                {isCurrentlyCuda && platform.metadata.isTauri && (
+                  <div className="space-y-3">
+                    <p className="text-sm text-muted-foreground">
+                      Running with CUDA GPU acceleration. Switch back to CPU if needed (you can
+                      re-download later).
+                    </p>
+                    <Button
+                      onClick={handleSwitchToCpu}
+                      variant="outline"
+                      className="w-full"
+                      size="sm"
+                    >
+                      <RotateCw className="h-4 w-4 mr-2" />
+                      Switch to CPU Backend
+                    </Button>
+                  </div>
+                )}
+
+                {/* Delete option when downloaded (and not active) */}
+                {cudaAvailable && !isCurrentlyCuda && (
+                  <Button
+                    onClick={handleDelete}
+                    variant="ghost"
+                    className="w-full text-muted-foreground hover:text-destructive"
+                    size="sm"
+                  >
+                    <Trash2 className="h-4 w-4 mr-2" />
+                    Remove CUDA Backend
+                  </Button>
+                )}
+              </div>
+            )}
+          </>
+        )}
+      </CardContent>
+    </Card>
+  );
+}

app/src/components/ServerSettings/ModelManagement.tsx

@@ -1,5 +1,5 @@
 import { useMutation, useQuery, useQueryClient } from '@tanstack/react-query';
-import { Download, Loader2, Trash2 } from 'lucide-react';
+import { ChevronDown, ChevronUp, Download, Loader2, RotateCcw, Trash2, X } from 'lucide-react';
 import { useCallback, useState } from 'react';
 import {
   AlertDialog,
@@ -16,6 +16,7 @@ import { Button } from '@/components/ui/button';
 import { Card, CardContent, CardDescription, CardHeader, CardTitle } from '@/components/ui/card';
 import { useToast } from '@/components/ui/use-toast';
 import { apiClient } from '@/lib/api/client';
+import type { ActiveDownloadTask } from '@/lib/api/types';
 import { useModelDownloadToast } from '@/lib/hooks/useModelDownloadToast';
 
 export function ModelManagement() {
@@ -23,6 +24,9 @@ export function ModelManagement() {
   const queryClient = useQueryClient();
   const [downloadingModel, setDownloadingModel] = useState<string | null>(null);
   const [downloadingDisplayName, setDownloadingDisplayName] = useState<string | null>(null);
+  const [consoleOpen, setConsoleOpen] = useState(false);
+  const [dismissedErrors, setDismissedErrors] = useState<Set<string>>(new Set());
+  const [localErrors, setLocalErrors] = useState<Map<string, string>>(new Map());
 
   const { data: modelStatus, isLoading } = useQuery({
     queryKey: ['modelStatus'],
@@ -35,19 +39,60 @@ export function ModelManagement() {
     refetchInterval: 5000, // Refresh every 5 seconds
   });
 
+  const { data: activeTasks } = useQuery({
+    queryKey: ['activeTasks'],
+    queryFn: () => apiClient.getActiveTasks(),
+    refetchInterval: 5000,
+  });
+
+  // Build a map of errored downloads for quick lookup, excluding dismissed ones
+  // Merge server errors with locally captured SSE errors
+  const erroredDownloads = new Map<string, ActiveDownloadTask>();
+  if (activeTasks?.downloads) {
+    for (const dl of activeTasks.downloads) {
+      if (dl.status === 'error' && !dismissedErrors.has(dl.model_name)) {
+        // Prefer locally captured error (from SSE) over server error
+        const localErr = localErrors.get(dl.model_name);
+        erroredDownloads.set(dl.model_name, localErr ? { ...dl, error: localErr } : dl);
+      }
+    }
+  }
+  // Also add locally captured errors that aren't in server response yet
+  for (const [modelName, error] of localErrors) {
+    if (!erroredDownloads.has(modelName) && !dismissedErrors.has(modelName)) {
+      erroredDownloads.set(modelName, {
+        model_name: modelName,
+        status: 'error',
+        started_at: new Date().toISOString(),
+        error,
+      });
+    }
+  }
+
+  const errorCount = erroredDownloads.size;
+
   // Callbacks for download completion
   const handleDownloadComplete = useCallback(() => {
     console.log('[ModelManagement] Download complete, clearing state');
     setDownloadingModel(null);
     setDownloadingDisplayName(null);
     queryClient.invalidateQueries({ queryKey: ['modelStatus'] });
+    queryClient.invalidateQueries({ queryKey: ['activeTasks'] });
   }, [queryClient]);
 
-  const handleDownloadError = useCallback(() => {
-    console.log('[ModelManagement] Download error, clearing state');
-    setDownloadingModel(null);
-    setDownloadingDisplayName(null);
-  }, []);
+  const handleDownloadError = useCallback(
+    (error: string) => {
+      console.log('[ModelManagement] Download error, clearing state');
+      if (downloadingModel) {
+        setLocalErrors((prev) => new Map(prev).set(downloadingModel, error));
+        setConsoleOpen(true);
+      }
+      setDownloadingModel(null);
+      setDownloadingDisplayName(null);
+      queryClient.invalidateQueries({ queryKey: ['activeTasks'] });
+    },
+    [queryClient, downloadingModel],
+  );
 
   // Use progress toast hook for the downloading model
   useModelDownloadToast({
@@ -67,11 +112,17 @@ export function ModelManagement() {
 
   const handleDownload = async (modelName: string) => {
     console.log('[Download] Button clicked for:', modelName, 'at', new Date().toISOString());
-    
+    // Clear any previous dismissal so fresh errors can appear
+    setDismissedErrors((prev) => {
+      const next = new Set(prev);
+      next.delete(modelName);
+      return next;
+    });
+
     // Find display name
     const model = modelStatus?.models.find((m) => m.model_name === modelName);
     const displayName = model?.display_name || modelName;
-    
+
     try {
       // IMPORTANT: Call the API FIRST before setting state
       // Setting state enables the SSE EventSource in useModelDownloadToast,
@@ -79,14 +130,15 @@ export function ModelManagement() {
       console.log('[Download] Calling download API for:', modelName);
       const result = await apiClient.triggerModelDownload(modelName);
       console.log('[Download] Download API responded:', result);
-      
+
       // NOW set state to enable SSE tracking (after download has started on backend)
       setDownloadingModel(modelName);
       setDownloadingDisplayName(displayName);
-      
+
       // Download initiated successfully - state will be cleared when SSE reports completion
       // or by the polling interval detecting the model is downloaded
       queryClient.invalidateQueries({ queryKey: ['modelStatus'] });
+      queryClient.invalidateQueries({ queryKey: ['activeTasks'] });
     } catch (error) {
       console.error('[Download] Download failed:', error);
       setDownloadingModel(null);
@@ -99,6 +151,61 @@ export function ModelManagement() {
     }
   };
 
+  const cancelMutation = useMutation({
+    mutationFn: (modelName: string) => apiClient.cancelDownload(modelName),
+    onSuccess: async () => {
+      await queryClient.invalidateQueries({ queryKey: ['modelStatus'], refetchType: 'all' });
+      await queryClient.invalidateQueries({ queryKey: ['activeTasks'], refetchType: 'all' });
+    },
+  });
+
+  const handleCancel = (modelName: string) => {
+    // Snapshot previous state for rollback
+    const prevDismissed = dismissedErrors;
+    const prevLocalErrors = localErrors;
+    const prevDownloadingModel = downloadingModel;
+    const prevDownloadingDisplayName = downloadingDisplayName;
+
+    // Optimistically hide the error and suppress downloading state in UI
+    setDismissedErrors((prev) => new Set(prev).add(modelName));
+    setLocalErrors((prev) => {
+      const next = new Map(prev);
+      next.delete(modelName);
+      return next;
+    });
+    if (downloadingModel === modelName) {
+      setDownloadingModel(null);
+      setDownloadingDisplayName(null);
+    }
+
+    cancelMutation.mutate(modelName, {
+      onError: () => {
+        // Rollback optimistic updates on failure
+        setDismissedErrors(prevDismissed);
+        setLocalErrors(prevLocalErrors);
+        setDownloadingModel(prevDownloadingModel);
+        setDownloadingDisplayName(prevDownloadingDisplayName);
+        toast({
+          title: 'Cancel failed',
+          description: 'Could not cancel the download task.',
+          variant: 'destructive',
+        });
+      },
+    });
+  };
+
+  const clearAllMutation = useMutation({
+    mutationFn: () => apiClient.clearAllTasks(),
+    onSuccess: async () => {
+      setDismissedErrors(new Set());
+      setLocalErrors(new Map());
+      setDownloadingModel(null);
+      setDownloadingDisplayName(null);
+      await queryClient.invalidateQueries({ queryKey: ['modelStatus'], refetchType: 'all' });
+      await queryClient.invalidateQueries({ queryKey: ['activeTasks'], refetchType: 'all' });
+    },
+  });
+
   const deleteMutation = useMutation({
     mutationFn: async (modelName: string) => {
       console.log('[Delete] Deleting model:', modelName);
@@ -114,14 +221,11 @@ export function ModelManagement() {
       });
       setDeleteDialogOpen(false);
       setModelToDelete(null);
-      // Invalidate AND explicitly refetch to ensure UI updates
-      // Using refetchType: 'all' ensures we refetch even if the query is stale
       console.log('[Delete] Invalidating modelStatus query');
-      await queryClient.invalidateQueries({ 
+      await queryClient.invalidateQueries({
         queryKey: ['modelStatus'],
         refetchType: 'all',
       });
-      // Also explicitly refetch to guarantee fresh data
       console.log('[Delete] Explicitly refetching modelStatus query');
       await queryClient.refetchQueries({ queryKey: ['modelStatus'] });
       console.log('[Delete] Query refetched');
@@ -178,13 +282,47 @@ export function ModelManagement() {
                         });
                         setDeleteDialogOpen(true);
                       }}
+                      onCancel={() => handleCancel(model.model_name)}
                       isDownloading={downloadingModel === model.model_name}
+                      isCancelling={
+                        cancelMutation.isPending && cancelMutation.variables === model.model_name
+                      }
+                      isDismissed={dismissedErrors.has(model.model_name)}
+                      erroredDownload={erroredDownloads.get(model.model_name)}
                       formatSize={formatSize}
                     />
                   ))}
               </div>
             </div>
 
+            {/* LuxTTS Models */}
+            {modelStatus.models.some((m) => m.model_name.startsWith('luxtts')) && (
+              <div>
+                <h3 className="text-sm font-semibold mb-3 text-muted-foreground">LuxTTS Models</h3>
+                <div className="space-y-2">
+                  {modelStatus.models
+                    .filter((m) => m.model_name.startsWith('luxtts'))
+                    .map((model) => (
+                      <ModelItem
+                        key={model.model_name}
+                        model={model}
+                        onDownload={() => handleDownload(model.model_name)}
+                        onDelete={() => {
+                          setModelToDelete({
+                            name: model.model_name,
+                            displayName: model.display_name,
+                            sizeMb: model.size_mb,
+                          });
+                          setDeleteDialogOpen(true);
+                        }}
+                        isDownloading={downloadingModel === model.model_name}
+                        formatSize={formatSize}
+                      />
+                    ))}
+                </div>
+              </div>
+            )}
+
             {/* Whisper Models */}
             <div>
               <h3 className="text-sm font-semibold mb-3 text-muted-foreground">
@@ -206,13 +344,79 @@ export function ModelManagement() {
                         });
                         setDeleteDialogOpen(true);
                       }}
+                      onCancel={() => handleCancel(model.model_name)}
                       isDownloading={downloadingModel === model.model_name}
+                      isCancelling={
+                        cancelMutation.isPending && cancelMutation.variables === model.model_name
+                      }
+                      isDismissed={dismissedErrors.has(model.model_name)}
+                      erroredDownload={erroredDownloads.get(model.model_name)}
                       formatSize={formatSize}
                     />
                   ))}
               </div>
             </div>
 
+            {/* Console Panel */}
+            {errorCount > 0 && (
+              <div className="border rounded-lg overflow-hidden">
+                <div className="flex items-center justify-between px-3 py-1.5 bg-muted/50 text-xs font-medium text-muted-foreground">
+                  <button
+                    type="button"
+                    onClick={() => setConsoleOpen((v) => !v)}
+                    className="flex items-center gap-2 hover:text-foreground transition-colors"
+                  >
+                    {consoleOpen ? (
+                      <ChevronUp className="h-3.5 w-3.5" />
+                    ) : (
+                      <ChevronDown className="h-3.5 w-3.5" />
+                    )}
+                    <span>Problems</span>
+                    <Badge variant="destructive" className="text-[10px] h-4 px-1.5 rounded-full">
+                      {errorCount}
+                    </Badge>
+                  </button>
+                  <Button
+                    size="sm"
+                    variant="ghost"
+                    className="h-6 px-2 text-xs text-muted-foreground hover:text-foreground"
+                    onClick={() => clearAllMutation.mutate()}
+                    disabled={clearAllMutation.isPending}
+                  >
+                    <RotateCcw className="h-3 w-3 mr-1" />
+                    Clear All
+                  </Button>
+                </div>
+                {consoleOpen && (
+                  <div className="bg-[#1e1e1e] text-[#d4d4d4] p-3 max-h-48 overflow-auto font-mono text-xs leading-relaxed">
+                    {Array.from(erroredDownloads.entries()).map(([modelName, dl]) => (
+                      <div key={modelName} className="mb-2 last:mb-0">
+                        <span className="text-[#f44747]">[error]</span>{' '}
+                        <span className="text-[#569cd6]">{modelName}</span>
+                        {dl.error ? (
+                          <>
+                            {': '}
+                            <span className="text-[#ce9178] whitespace-pre-wrap break-all">
+                              {dl.error}
+                            </span>
+                          </>
+                        ) : (
+                          <>
+                            {': '}
+                            <span className="text-[#808080]">
+                              No error details available. Try downloading again.
+                            </span>
+                          </>
+                        )}
+                        <div className="text-[#6a9955] mt-0.5">
+                          started at {new Date(dl.started_at).toLocaleString()}
+                        </div>
+                      </div>
+                    ))}
+                  </div>
+                )}
+              </div>
+            )}
           </div>
         ) : null}
       </CardContent>
@@ -265,23 +469,38 @@ interface ModelItemProps {
     model_name: string;
     display_name: string;
     downloaded: boolean;
-    downloading?: boolean;  // From server - true if download in progress
+    downloading?: boolean; // From server - true if download in progress
     size_mb?: number;
     loaded: boolean;
   };
   onDownload: () => void;
   onDelete: () => void;
-  isDownloading: boolean;  // Local state - true if user just clicked download
+  onCancel: () => void;
+  isDownloading: boolean; // Local state - true if user just clicked download
+  isCancelling: boolean;
+  isDismissed: boolean;
+  erroredDownload?: ActiveDownloadTask;
   formatSize: (sizeMb?: number) => string;
 }
 
-function ModelItem({ model, onDownload, onDelete, isDownloading, formatSize }: ModelItemProps) {
+function ModelItem({
+  model,
+  onDownload,
+  onDelete,
+  onCancel,
+  isDownloading,
+  isCancelling,
+  isDismissed,
+  erroredDownload,
+  formatSize,
+}: ModelItemProps) {
   // Use server's downloading state OR local state (for immediate feedback before server updates)
-  const showDownloading = model.downloading || isDownloading;
-  
+  // Suppress downloading if user just dismissed/cancelled this model
+  const showDownloading = (model.downloading || isDownloading) && !erroredDownload && !isDismissed;
+
   return (
     <div className="flex items-center justify-between p-3 border rounded-lg">
-      <div className="flex-1">
+      <div className="flex-1 min-w-0">
         <div className="flex items-center gap-2">
           <span className="font-medium text-sm">{model.display_name}</span>
           {model.loaded && (
@@ -289,21 +508,41 @@ function ModelItem({ model, onDownload, onDelete, isDownloading, formatSize }: M
               Loaded
             </Badge>
           )}
-          {/* Only show Downloaded if actually downloaded AND not downloading */}
-          {model.downloaded && !model.loaded && !showDownloading && (
+          {model.downloaded && !model.loaded && !showDownloading && !erroredDownload && (
             <Badge variant="secondary" className="text-xs">
               Downloaded
             </Badge>
           )}
+          {erroredDownload && (
+            <Badge variant="destructive" className="text-xs">
+              Error
+            </Badge>
+          )}
         </div>
-        {model.downloaded && model.size_mb && !showDownloading && (
+        {model.downloaded && model.size_mb && !showDownloading && !erroredDownload && (
           <div className="text-xs text-muted-foreground mt-1">
             Size: {formatSize(model.size_mb)}
           </div>
         )}
       </div>
-      <div className="flex items-center gap-2">
-        {model.downloaded && !showDownloading ? (
+      <div className="flex items-center gap-2 shrink-0 ml-2">
+        {erroredDownload ? (
+          <div className="flex items-center gap-2">
+            <Button size="sm" onClick={onDownload} variant="outline">
+              <Download className="h-4 w-4 mr-2" />
+              Retry
+            </Button>
+            <Button
+              size="sm"
+              onClick={onCancel}
+              variant="ghost"
+              disabled={isCancelling}
+              title="Dismiss error"
+            >
+              <X className="h-4 w-4" />
+            </Button>
+          </div>
+        ) : model.downloaded && !showDownloading ? (
           <div className="flex items-center gap-2">
             <div className="flex items-center gap-1 text-sm text-muted-foreground">
               <span>Ready</span>
@@ -319,10 +558,21 @@ function ModelItem({ model, onDownload, onDelete, isDownloading, formatSize }: M
             </Button>
           </div>
         ) : showDownloading ? (
-          <Button size="sm" variant="outline" disabled>
-            <Loader2 className="h-4 w-4 mr-2 animate-spin" />
-            Downloading...
-          </Button>
+          <div className="flex items-center gap-2">
+            <Button size="sm" variant="outline" disabled>
+              <Loader2 className="h-4 w-4 mr-2 animate-spin" />
+              Downloading...
+            </Button>
+            <Button
+              size="sm"
+              onClick={onCancel}
+              variant="ghost"
+              disabled={isCancelling}
+              title="Cancel download"
+            >
+              <X className="h-4 w-4" />
+            </Button>
+          </div>
         ) : (
           <Button size="sm" onClick={onDownload} variant="outline">
             <Download className="h-4 w-4 mr-2" />

app/src/components/ServerTab/ServerTab.tsx

@@ -1,4 +1,5 @@
 import { ConnectionForm } from '@/components/ServerSettings/ConnectionForm';
+import { GpuAcceleration } from '@/components/ServerSettings/GpuAcceleration';
 import { ServerStatus } from '@/components/ServerSettings/ServerStatus';
 import { UpdateStatus } from '@/components/ServerSettings/UpdateStatus';
 import { usePlatform } from '@/platform/PlatformContext';
@@ -11,6 +12,7 @@ export function ServerTab() {
         <ConnectionForm />
         <ServerStatus />
       </div>
+      {platform.metadata.isTauri && <GpuAcceleration />}
       {platform.metadata.isTauri && <UpdateStatus />}
       <div className="py-8 text-center text-sm text-muted-foreground">
         Created by{' '}

app/src/components/VoiceProfiles/AudioSampleRecording.tsx

@@ -58,6 +58,7 @@ export function AudioSampleRecording({
   // Request microphone access when component mounts
   useEffect(() => {
     if (!showWaveform) return;
+    if (!navigator.mediaDevices || !navigator.mediaDevices.getUserMedia) return;
 
     let stream: MediaStream | null = null;
 

app/src/components/VoiceProfiles/ProfileForm.tsx

@@ -43,7 +43,7 @@ import {
 } from '@/lib/hooks/useProfiles';
 import { useSystemAudioCapture } from '@/lib/hooks/useSystemAudioCapture';
 import { useTranscription } from '@/lib/hooks/useTranscription';
-import { formatAudioDuration, getAudioDuration } from '@/lib/utils/audio';
+import { convertToWav, formatAudioDuration, getAudioDuration } from '@/lib/utils/audio';
 import { usePlatform } from '@/platform/PlatformContext';
 import { useServerStore } from '@/stores/serverStore';
 import { type ProfileFormDraft, useUIStore } from '@/stores/uiStore';
@@ -505,10 +505,23 @@ export function ProfileForm() {
           language: data.language,
         });
 
+        // Convert non-WAV uploads to WAV so the backend can always use soundfile.
+        // Recorded audio is already WAV (from useAudioRecording's convertToWav call).
+        let fileToUpload: File = sampleFile;
+        if (!sampleFile.type.includes('wav') && !sampleFile.name.toLowerCase().endsWith('.wav')) {
+          try {
+            const wavBlob = await convertToWav(sampleFile);
+            const wavName = sampleFile.name.replace(/\.[^.]+$/, '.wav');
+            fileToUpload = new File([wavBlob], wavName, { type: 'audio/wav' });
+          } catch {
+            // If browser can't decode the format, send the original and let the backend try.
+          }
+        }
+
         try {
           await addSample.mutateAsync({
             profileId: profile.id,
-            file: sampleFile,
+            file: fileToUpload,
             referenceText: referenceText,
           });
 

app/src/components/VoicesTab/VoicesTab.tsx

@@ -79,8 +79,8 @@ export function VoicesTab() {
     setDialogOpen(true);
   };
 
-  const handleDelete = (profileId: string) => {
-    if (confirm('Are you sure you want to delete this profile?')) {
+  const handleProfileDelete = async (profileId: string) => {
+    if (await confirm('Are you sure you want to delete this profile?')) {
       deleteProfile.mutate(profileId);
     }
   };
@@ -147,7 +147,7 @@ export function VoicesTab() {
                 channels={channels || []}
                 onChannelChange={(channelIds) => handleChannelChange(profile.id, channelIds)}
                 onEdit={() => handleEdit(profile.id)}
-                onDelete={() => handleDelete(profile.id)}
+                onDelete={() => handleProfileDelete(profile.id)}
               />
             ))}
           </TableBody>

app/src/lib/api/client.ts

@@ -1,29 +1,30 @@
-import { useServerStore } from '@/stores/serverStore';
 import type { LanguageCode } from '@/lib/constants/languages';
+import { useServerStore } from '@/stores/serverStore';
 import type {
-  VoiceProfileCreate,
-  VoiceProfileResponse,
-  ProfileSampleResponse,
+  ActiveTasksResponse,
+  CudaStatus,
   GenerationRequest,
   GenerationResponse,
-  HistoryQuery,
-  HistoryListResponse,
-  HistoryResponse,
-  TranscriptionResponse,
   HealthResponse,
-  ModelStatusListResponse,
+  HistoryListResponse,
+  HistoryQuery,
+  HistoryResponse,
   ModelDownloadRequest,
-  ActiveTasksResponse,
+  ModelStatusListResponse,
+  ProfileSampleResponse,
   StoryCreate,
-  StoryResponse,
   StoryDetailResponse,
+  StoryItemBatchUpdate,
   StoryItemCreate,
   StoryItemDetail,
-  StoryItemBatchUpdate,
-  StoryItemReorder,
   StoryItemMove,
-  StoryItemTrim,
+  StoryItemReorder,
   StoryItemSplit,
+  StoryItemTrim,
+  StoryResponse,
+  TranscriptionResponse,
+  VoiceProfileCreate,
+  VoiceProfileResponse,
 } from './types';
 
 class ApiClient {
@@ -251,7 +252,13 @@ class ApiClient {
     return response.blob();
   }
 
-  async importGeneration(file: File): Promise<{ id: string; profile_id: string; profile_name: string; text: string; message: string }> {
+  async importGeneration(file: File): Promise<{
+    id: string;
+    profile_id: string;
+    profile_name: string;
+    text: string;
+    message: string;
+  }> {
     const url = `${this.getBaseUrl()}/history/import`;
     const formData = new FormData();
     formData.append('file', file);
@@ -310,7 +317,12 @@ class ApiClient {
   }
 
   async triggerModelDownload(modelName: string): Promise<{ message: string }> {
-    console.log('[API] triggerModelDownload called for:', modelName, 'at', new Date().toISOString());
+    console.log(
+      '[API] triggerModelDownload called for:',
+      modelName,
+      'at',
+      new Date().toISOString(),
+    );
     const result = await this.request<{ message: string }>('/models/download', {
       method: 'POST',
       body: JSON.stringify({ model_name: modelName } as ModelDownloadRequest),
@@ -325,11 +337,22 @@ class ApiClient {
     });
   }
 
+  async cancelDownload(modelName: string): Promise<{ message: string }> {
+    return this.request<{ message: string }>('/models/download/cancel', {
+      method: 'POST',
+      body: JSON.stringify({ model_name: modelName } as ModelDownloadRequest),
+    });
+  }
+
   // Task Management
   async getActiveTasks(): Promise<ActiveTasksResponse> {
     return this.request<ActiveTasksResponse>('/tasks/active');
   }
 
+  async clearAllTasks(): Promise<{ message: string }> {
+    return this.request<{ message: string }>('/tasks/clear', { method: 'POST' });
+  }
+
   // Audio Channels
   async listChannels(): Promise<
     Array<{
@@ -343,10 +366,7 @@ class ApiClient {
     return this.request('/channels');
   }
 
-  async createChannel(data: {
-    name: string;
-    device_ids: string[];
-  }): Promise<{
+  async createChannel(data: { name: string; device_ids: string[] }): Promise<{
     id: string;
     name: string;
     is_default: boolean;
@@ -388,10 +408,7 @@ class ApiClient {
     return this.request(`/channels/${channelId}/voices`);
   }
 
-  async setChannelVoices(
-    channelId: string,
-    profileIds: string[],
-  ): Promise<{ message: string }> {
+  async setChannelVoices(channelId: string, profileIds: string[]): Promise<{ message: string }> {
     return this.request(`/channels/${channelId}/voices`, {
       method: 'PUT',
       body: JSON.stringify({ profile_ids: profileIds }),
@@ -402,16 +419,30 @@ class ApiClient {
     return this.request(`/profiles/${profileId}/channels`);
   }
 
-  async setProfileChannels(
-    profileId: string,
-    channelIds: string[],
-  ): Promise<{ message: string }> {
+  async setProfileChannels(profileId: string, channelIds: string[]): Promise<{ message: string }> {
     return this.request(`/profiles/${profileId}/channels`, {
       method: 'PUT',
       body: JSON.stringify({ channel_ids: channelIds }),
     });
   }
 
+  // CUDA Backend Management
+  async getCudaStatus(): Promise<CudaStatus> {
+    return this.request<CudaStatus>('/backend/cuda-status');
+  }
+
+  async downloadCudaBackend(): Promise<{ message: string; progress_key: string }> {
+    return this.request<{ message: string; progress_key: string }>('/backend/download-cuda', {
+      method: 'POST',
+    });
+  }
+
+  async deleteCudaBackend(): Promise<{ message: string }> {
+    return this.request<{ message: string }>('/backend/cuda', {
+      method: 'DELETE',
+    });
+  }
+
   // Stories
   async listStories(): Promise<StoryResponse[]> {
     return this.request<StoryResponse[]>('/stories');
@@ -468,21 +499,33 @@ class ApiClient {
     });
   }
 
-  async moveStoryItem(storyId: string, itemId: string, data: StoryItemMove): Promise<StoryItemDetail> {
+  async moveStoryItem(
+    storyId: string,
+    itemId: string,
+    data: StoryItemMove,
+  ): Promise<StoryItemDetail> {
     return this.request<StoryItemDetail>(`/stories/${storyId}/items/${itemId}/move`, {
       method: 'PUT',
       body: JSON.stringify(data),
     });
   }
 
-  async trimStoryItem(storyId: string, itemId: string, data: StoryItemTrim): Promise<StoryItemDetail> {
+  async trimStoryItem(
+    storyId: string,
+    itemId: string,
+    data: StoryItemTrim,
+  ): Promise<StoryItemDetail> {
     return this.request<StoryItemDetail>(`/stories/${storyId}/items/${itemId}/trim`, {
       method: 'PUT',
       body: JSON.stringify(data),
     });
   }
 
-  async splitStoryItem(storyId: string, itemId: string, data: StoryItemSplit): Promise<StoryItemDetail[]> {
+  async splitStoryItem(
+    storyId: string,
+    itemId: string,
+    data: StoryItemSplit,
+  ): Promise<StoryItemDetail[]> {
     return this.request<StoryItemDetail[]>(`/stories/${storyId}/items/${itemId}/split`, {
       method: 'POST',
       body: JSON.stringify(data),

app/src/lib/api/types.ts

@@ -34,6 +34,8 @@ export interface GenerationRequest {
   language: LanguageCode;
   seed?: number;
   model_size?: '1.7B' | '0.6B';
+  engine?: 'qwen' | 'luxtts';
+  instruct?: string;
 }
 
 export interface GenerationResponse {
@@ -78,7 +80,29 @@ export interface HealthResponse {
   model_downloaded?: boolean;
   model_size?: string;
   gpu_available: boolean;
+  gpu_type?: string;
   vram_used_mb?: number;
+  backend_type?: string;
+  backend_variant?: string; // "cpu" or "cuda"
+}
+
+export interface CudaDownloadProgress {
+  model_name: string;
+  current: number;
+  total: number;
+  progress: number;
+  filename?: string;
+  status: 'downloading' | 'extracting' | 'complete' | 'error';
+  timestamp: string;
+  error?: string;
+}
+
+export interface CudaStatus {
+  available: boolean; // CUDA binary exists on disk
+  active: boolean; // Currently running the CUDA binary
+  binary_path?: string;
+  downloading: boolean; // Download in progress
+  download_progress?: CudaDownloadProgress;
 }
 
 export interface ModelProgress {
@@ -96,7 +120,7 @@ export interface ModelStatus {
   model_name: string;
   display_name: string;
   downloaded: boolean;
-  downloading: boolean;  // True if download is in progress
+  downloading: boolean; // True if download is in progress
   size_mb?: number;
   loaded: boolean;
 }
@@ -113,6 +137,7 @@ export interface ActiveDownloadTask {
   model_name: string;
   status: string;
   started_at: string;
+  error?: string;
 }
 
 export interface ActiveGenerationTask {

app/src/lib/hooks/useAudioRecording.ts

@@ -20,11 +20,13 @@ export function useAudioRecording({
   const streamRef = useRef<MediaStream | null>(null);
   const timerRef = useRef<number | null>(null);
   const startTimeRef = useRef<number | null>(null);
+  const cancelledRef = useRef<boolean>(false);
 
   const startRecording = useCallback(async () => {
     try {
       setError(null);
       chunksRef.current = [];
+      cancelledRef.current = false;
       setDuration(0);
 
       // Check if getUserMedia is available
@@ -87,31 +89,34 @@ export function useAudioRecording({
       };
 
       mediaRecorder.onstop = async () => {
+        // Snapshot the cancellation flag and recorded duration immediately —
+        // cancelRecording() clears chunks and sets cancelledRef synchronously
+        // before this async handler runs, so we must check it first.
+        const wasCancelled = cancelledRef.current;
+        const recordedDuration = startTimeRef.current
+          ? (Date.now() - startTimeRef.current) / 1000
+          : undefined;
+
         const webmBlob = new Blob(chunksRef.current, { type: 'audio/webm' });
 
-        // Convert to WAV format to avoid needing ffmpeg on backend
-        try {
-          const wavBlob = await convertToWav(webmBlob);
-
-          // Pass the actual recorded duration
-          const recordedDuration = startTimeRef.current
-            ? (Date.now() - startTimeRef.current) / 1000
-            : undefined;
-          onRecordingComplete?.(wavBlob, recordedDuration);
-        } catch (err) {
-          console.error('Error converting audio to WAV:', err);
-          // Fallback to original blob if conversion fails
-          const recordedDuration = startTimeRef.current
-            ? (Date.now() - startTimeRef.current) / 1000
-            : undefined;
-          onRecordingComplete?.(webmBlob, recordedDuration);
-        }
-
-        // Stop all tracks
+        // Stop all tracks now that we have the data
         streamRef.current?.getTracks().forEach((track) => {
           track.stop();
         });
         streamRef.current = null;
+
+        // Don't fire completion callback if the recording was cancelled
+        if (wasCancelled) return;
+
+        // Convert to WAV format to avoid needing ffmpeg on backend
+        try {
+          const wavBlob = await convertToWav(webmBlob);
+          onRecordingComplete?.(wavBlob, recordedDuration);
+        } catch (err) {
+          console.error('Error converting audio to WAV:', err);
+          // Fallback to original blob if conversion fails
+          onRecordingComplete?.(webmBlob, recordedDuration);
+        }
       };
 
       mediaRecorder.onerror = (event) => {
@@ -167,9 +172,10 @@ export function useAudioRecording({
 
   const cancelRecording = useCallback(() => {
     if (mediaRecorderRef.current) {
+      cancelledRef.current = true; // Must be set before stop() triggers onstop
+      chunksRef.current = [];
       mediaRecorderRef.current.stop();
       setIsRecording(false);
-      chunksRef.current = [];
       setDuration(0);
     }
 

app/src/lib/hooks/useGenerationForm.ts

@@ -16,6 +16,7 @@ const generationSchema = z.object({
   seed: z.number().int().optional(),
   modelSize: z.enum(['1.7B', '0.6B']).optional(),
   instruct: z.string().max(500).optional(),
+  engine: z.enum(['qwen', 'luxtts']).optional(),
 });
 
 export type GenerationFormValues = z.infer<typeof generationSchema>;
@@ -47,6 +48,7 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
       seed: undefined,
       modelSize: '1.7B',
       instruct: '',
+      engine: 'qwen',
       ...options.defaultValues,
     },
   });
@@ -67,8 +69,14 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
     try {
       setIsGenerating(true);
 
-      const modelName = `qwen-tts-${data.modelSize}`;
-      const displayName = data.modelSize === '1.7B' ? 'Qwen TTS 1.7B' : 'Qwen TTS 0.6B';
+      const engine = data.engine || 'qwen';
+      const modelName = engine === 'luxtts' ? 'luxtts' : `qwen-tts-${data.modelSize}`;
+      const displayName =
+        engine === 'luxtts'
+          ? 'LuxTTS'
+          : data.modelSize === '1.7B'
+            ? 'Qwen TTS 1.7B'
+            : 'Qwen TTS 0.6B';
 
       try {
         const modelStatus = await apiClient.getModelStatus();
@@ -87,8 +95,9 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
         text: data.text,
         language: data.language,
         seed: data.seed,
-        model_size: data.modelSize,
-        instruct: data.instruct || undefined,
+        model_size: engine === 'luxtts' ? undefined : data.modelSize,
+        engine,
+        instruct: engine === 'luxtts' ? undefined : data.instruct || undefined,
       });
 
       toast({
@@ -99,7 +108,14 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
       const audioUrl = apiClient.getAudioUrl(result.id);
       setAudioWithAutoPlay(audioUrl, result.id, selectedProfileId, data.text.substring(0, 50));
 
-      form.reset();
+      form.reset({
+        text: '',
+        language: data.language,
+        seed: undefined,
+        modelSize: data.modelSize,
+        instruct: '',
+        engine: data.engine,
+      });
       options.onSuccess?.(result.id);
     } catch (error) {
       toast({

app/src/lib/hooks/useModelDownloadToast.tsx

@@ -10,7 +10,7 @@ interface UseModelDownloadToastOptions {
   displayName: string;
   enabled?: boolean;
   onComplete?: () => void;
-  onError?: () => void;
+  onError?: (error: string) => void;
 }
 
 /**
@@ -101,7 +101,7 @@ export function useModelDownloadToast({
               break;
             case 'error':
               statusIcon = <XCircle className="h-4 w-4 text-destructive" />;
-              statusText = `Error: ${progress.error || 'Unknown error'}`;
+              statusText = 'Download failed. See Problems panel for details.';
               break;
             case 'downloading':
               statusIcon = <Loader2 className="h-4 w-4 animate-spin" />;
@@ -131,8 +131,7 @@ export function useModelDownloadToast({
                 )}
               </div>
             ),
-            duration: progress.status === 'complete' ? 5000 : Infinity,
-            variant: progress.status === 'error' ? 'destructive' : 'default',
+            duration: progress.status === 'complete' || progress.status === 'error' ? 5000 : Infinity,
           });
 
           // Close connection and dismiss toast on completion or error
@@ -169,7 +168,7 @@ export function useModelDownloadToast({
               onComplete();
             } else if (isError && onError) {
               console.log('[useModelDownloadToast] Download error, calling onError callback');
-              onError();
+              onError(progress.error || 'Unknown error');
             }
           }
         }

app/src/lib/utils/audio.ts

@@ -22,6 +22,11 @@ export function formatAudioDuration(seconds: number): string {
  * If the file has a recordedDuration property (from recording hooks),
  * use that instead of trying to read metadata. This fixes issues on Windows
  * where WebM files from MediaRecorder don't have proper duration metadata.
+ *
+ * For uploaded files we use AudioContext.decodeAudioData which fully decodes
+ * the audio and returns the exact duration. This is more reliable than
+ * HTMLMediaElement.duration which can return incorrect large values for VBR
+ * MP3 files that lack a proper XING/VBRI header.
  */
 export async function getAudioDuration(
   file: File & { recordedDuration?: number },
@@ -30,26 +35,39 @@ export async function getAudioDuration(
     return file.recordedDuration;
   }
 
-  return new Promise((resolve, reject) => {
-    const audio = new Audio();
-    const url = URL.createObjectURL(file);
+  // Use Web Audio API for accurate duration — avoids VBR MP3 metadata issues.
+  try {
+    const audioContext = new AudioContext();
+    try {
+      const arrayBuffer = await file.arrayBuffer();
+      const audioBuffer = await audioContext.decodeAudioData(arrayBuffer);
+      return audioBuffer.duration;
+    } finally {
+      await audioContext.close();
+    }
+  } catch {
+    // Fallback: read duration from the media element (less accurate but works for WAV).
+    return new Promise((resolve, reject) => {
+      const audio = new Audio();
+      const url = URL.createObjectURL(file);
 
-    audio.addEventListener('loadedmetadata', () => {
-      URL.revokeObjectURL(url);
-      if (Number.isFinite(audio.duration) && audio.duration > 0) {
-        resolve(audio.duration);
-      } else {
-        reject(new Error('Audio file has invalid duration metadata'));
-      }
+      audio.addEventListener('loadedmetadata', () => {
+        URL.revokeObjectURL(url);
+        if (Number.isFinite(audio.duration) && audio.duration > 0) {
+          resolve(audio.duration);
+        } else {
+          reject(new Error('Audio file has invalid duration metadata'));
+        }
+      });
+
+      audio.addEventListener('error', () => {
+        URL.revokeObjectURL(url);
+        reject(new Error('Failed to load audio file'));
+      });
+
+      audio.src = url;
     });
-
-    audio.addEventListener('error', () => {
-      URL.revokeObjectURL(url);
-      reject(new Error('Failed to load audio file'));
-    });
-
-    audio.src = url;
-  });
+  }
 }
 
 /**

app/src/platform/types.ts

@@ -51,6 +51,7 @@ export interface PlatformAudio {
 export interface PlatformLifecycle {
   startServer(remote?: boolean): Promise<string>;
   stopServer(): Promise<void>;
+  restartServer(): Promise<string>;
   setKeepServerRunning(keep: boolean): Promise<void>;
   setupWindowCloseHandler(): Promise<void>;
   onServerReady?: () => void;

backend/__init__.py

@@ -1,3 +1,3 @@
 # Backend package
 
-__version__ = "0.1.12"
+__version__ = "0.1.13"

backend/backends/__init__.py

@@ -4,6 +4,7 @@ Backend abstraction layer for TTS and STT.
 Provides a unified interface for MLX and PyTorch backends.
 """
 
+import threading
 from typing import Protocol, Optional, Tuple, List
 from typing_extensions import runtime_checkable
 import numpy as np
@@ -112,29 +113,65 @@ class STTBackend(Protocol):
 
 # Global backend instances
 _tts_backend: Optional[TTSBackend] = None
+_tts_backends: dict[str, TTSBackend] = {}
+_tts_backends_lock = threading.Lock()
 _stt_backend: Optional[STTBackend] = None
 
+# Supported TTS engines
+TTS_ENGINES = {
+    "qwen": "Qwen TTS",
+    "luxtts": "LuxTTS",
+}
+
 
 def get_tts_backend() -> TTSBackend:
     """
-    Get or create TTS backend instance based on platform.
+    Get or create the default (Qwen) TTS backend instance based on platform.
     
     Returns:
         TTS backend instance (MLX or PyTorch)
     """
-    global _tts_backend
+    return get_tts_backend_for_engine("qwen")
+
+
+def get_tts_backend_for_engine(engine: str) -> TTSBackend:
+    """
+    Get or create a TTS backend for the given engine.
     
-    if _tts_backend is None:
-        backend_type = get_backend_type()
+    Args:
+        engine: Engine name ("qwen" or "luxtts")
+    
+    Returns:
+        TTS backend instance
+    """
+    global _tts_backends
+    
+    # Fast path: check without lock
+    if engine in _tts_backends:
+        return _tts_backends[engine]
+    
+    # Slow path: create with lock to avoid duplicate instantiation
+    with _tts_backends_lock:
+        # Double-check after acquiring lock
+        if engine in _tts_backends:
+            return _tts_backends[engine]
         
-        if backend_type == "mlx":
-            from .mlx_backend import MLXTTSBackend
-            _tts_backend = MLXTTSBackend()
+        if engine == "qwen":
+            backend_type = get_backend_type()
+            if backend_type == "mlx":
+                from .mlx_backend import MLXTTSBackend
+                backend = MLXTTSBackend()
+            else:
+                from .pytorch_backend import PyTorchTTSBackend
+                backend = PyTorchTTSBackend()
+        elif engine == "luxtts":
+            from .luxtts_backend import LuxTTSBackend
+            backend = LuxTTSBackend()
         else:
-            from .pytorch_backend import PyTorchTTSBackend
-            _tts_backend = PyTorchTTSBackend()
-    
-    return _tts_backend
+            raise ValueError(f"Unknown TTS engine: {engine}. Supported: {list(TTS_ENGINES.keys())}")
+        
+        _tts_backends[engine] = backend
+        return backend
 
 
 def get_stt_backend() -> STTBackend:
@@ -161,6 +198,7 @@ def get_stt_backend() -> STTBackend:
 
 def reset_backends():
     """Reset backend instances (useful for testing)."""
-    global _tts_backend, _stt_backend
+    global _tts_backend, _tts_backends, _stt_backend
     _tts_backend = None
+    _tts_backends.clear()
     _stt_backend = None

backend/backends/luxtts_backend.py

@@ -0,0 +1,264 @@
+"""
+LuxTTS backend implementation.
+
+Wraps the LuxTTS (ZipVoice) model for zero-shot voice cloning.
+~1GB VRAM, 48kHz output, 150x realtime on CPU.
+"""
+
+import asyncio
+import logging
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from . import TTSBackend
+from ..utils.audio import normalize_audio, load_audio
+from ..utils.cache import get_cache_key, get_cached_voice_prompt, cache_voice_prompt
+from ..utils.progress import get_progress_manager
+from ..utils.tasks import get_task_manager
+
+logger = logging.getLogger(__name__)
+
+# HuggingFace repo for model weight detection
+LUXTTS_HF_REPO = "YatharthS/LuxTTS"
+
+
+class LuxTTSBackend:
+    """LuxTTS backend for zero-shot voice cloning."""
+
+    def __init__(self):
+        self.model = None
+        self.model_size = "default"  # LuxTTS has only one model size
+        self._device = None
+
+    def _get_device(self) -> str:
+        """Get the best available device."""
+        import torch
+
+        if torch.cuda.is_available():
+            return "cuda"
+        if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+            return "mps"
+        return "cpu"
+
+    def is_loaded(self) -> bool:
+        return self.model is not None
+
+    @property
+    def device(self) -> str:
+        if self._device is None:
+            self._device = self._get_device()
+        return self._device
+
+    def _get_model_path(self, model_size: str) -> str:
+        return LUXTTS_HF_REPO
+
+    def _is_model_cached(self, model_size: str = "default") -> bool:
+        """Check if LuxTTS model weights are cached locally."""
+        try:
+            from huggingface_hub import constants as hf_constants
+
+            repo_cache = (
+                Path(hf_constants.HF_HUB_CACHE)
+                / ("models--" + LUXTTS_HF_REPO.replace("/", "--"))
+            )
+
+            if not repo_cache.exists():
+                return False
+
+            blobs_dir = repo_cache / "blobs"
+            if blobs_dir.exists() and any(blobs_dir.glob("*.incomplete")):
+                return False
+
+            snapshots_dir = repo_cache / "snapshots"
+            if snapshots_dir.exists():
+                has_weights = any(snapshots_dir.rglob("*.pt")) or any(
+                    snapshots_dir.rglob("*.safetensors")
+                ) or any(snapshots_dir.rglob("*.onnx")) or any(
+                    snapshots_dir.rglob("*.bin")
+                )
+                return has_weights
+
+            return False
+        except Exception as e:
+            logger.warning(f"Error checking LuxTTS cache: {e}")
+            return False
+
+    async def load_model(self, model_size: str = "default") -> None:
+        """Load the LuxTTS model."""
+        if self.model is not None:
+            return
+
+        await asyncio.to_thread(self._load_model_sync)
+
+    def _load_model_sync(self):
+        """Synchronous model loading."""
+        progress_manager = get_progress_manager()
+        task_manager = get_task_manager()
+        model_name = "luxtts"
+
+        is_cached = self._is_model_cached()
+
+        if not is_cached:
+            task_manager.start_download(model_name)
+            progress_manager.update_progress(
+                model_name=model_name,
+                current=0,
+                total=0,
+                filename="Downloading LuxTTS model...",
+                status="downloading",
+            )
+
+        try:
+            from zipvoice.luxvoice import LuxTTS
+
+            device = self.device
+            logger.info(f"Loading LuxTTS on {device}...")
+
+            # LuxTTS constructor downloads model and loads everything
+            if device == "cpu":
+                import os
+                threads = os.cpu_count() or 4
+                self.model = LuxTTS(
+                    model_path=LUXTTS_HF_REPO,
+                    device="cpu",
+                    threads=min(threads, 8),
+                )
+            else:
+                self.model = LuxTTS(
+                    model_path=LUXTTS_HF_REPO,
+                    device=device,
+                )
+
+            if not is_cached:
+                progress_manager.mark_complete(model_name)
+                task_manager.complete_download(model_name)
+
+            logger.info("LuxTTS loaded successfully")
+
+        except Exception as e:
+            logger.error(f"Failed to load LuxTTS: {e}")
+            if not is_cached:
+                progress_manager.mark_error(model_name, str(e))
+                task_manager.error_download(model_name, str(e))
+            raise
+
+    def unload_model(self) -> None:
+        """Unload model to free memory."""
+        if self.model is not None:
+            del self.model
+            self.model = None
+
+            import torch
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+            logger.info("LuxTTS unloaded")
+
+    async def create_voice_prompt(
+        self,
+        audio_path: str,
+        reference_text: str,
+        use_cache: bool = True,
+    ) -> Tuple[dict, bool]:
+        """
+        Create voice prompt from reference audio.
+
+        LuxTTS uses its own encode_prompt() which runs Whisper ASR internally
+        to transcribe the reference. The reference_text parameter is not used
+        by LuxTTS itself, but we include it in the cache key for consistency.
+        """
+        await self.load_model()
+
+        # Compute cache key once for both lookup and storage
+        cache_key = ("luxtts_" + get_cache_key(audio_path, reference_text)) if use_cache else None
+
+        if cache_key:
+            cached = get_cached_voice_prompt(cache_key)
+            if cached is not None and isinstance(cached, dict):
+                return cached, True
+
+        def _encode_sync():
+            return self.model.encode_prompt(
+                prompt_audio=str(audio_path),
+                duration=5,
+                rms=0.01,
+            )
+
+        encoded = await asyncio.to_thread(_encode_sync)
+
+        if cache_key:
+            cache_voice_prompt(cache_key, encoded)
+
+        return encoded, False
+
+    async def combine_voice_prompts(
+        self,
+        audio_paths: List[str],
+        reference_texts: List[str],
+    ) -> Tuple[np.ndarray, str]:
+        """
+        Combine multiple reference samples.
+
+        LuxTTS doesn't have native multi-prompt support, so we concatenate
+        the audio and let encode_prompt handle the combined clip.
+        """
+        combined_audio = []
+        for path in audio_paths:
+            audio, _sr = load_audio(path, sample_rate=24000)
+            audio = normalize_audio(audio)
+            combined_audio.append(audio)
+
+        mixed = np.concatenate(combined_audio)
+        mixed = normalize_audio(mixed)
+        combined_text = " ".join(reference_texts)
+
+        return mixed, combined_text
+
+    async def generate(
+        self,
+        text: str,
+        voice_prompt: dict,
+        language: str = "en",
+        seed: Optional[int] = None,
+        instruct: Optional[str] = None,
+    ) -> Tuple[np.ndarray, int]:
+        """
+        Generate audio from text using LuxTTS.
+
+        Args:
+            text: Text to synthesize
+            voice_prompt: Encoded prompt dict from encode_prompt()
+            language: Language code (LuxTTS is English-focused)
+            seed: Random seed for reproducibility
+            instruct: Not supported by LuxTTS (ignored)
+
+        Returns:
+            Tuple of (audio_array, sample_rate)
+        """
+        await self.load_model()
+
+        def _generate_sync():
+            import torch
+
+            if seed is not None:
+                torch.manual_seed(seed)
+                if torch.cuda.is_available():
+                    torch.cuda.manual_seed(seed)
+
+            wav = self.model.generate_speech(
+                text=text,
+                encode_dict=voice_prompt,
+                num_steps=4,
+                guidance_scale=3.0,
+                t_shift=0.5,
+                speed=1.0,
+                return_smooth=False,  # 48kHz output
+            )
+
+            # LuxTTS returns a tensor (may be on GPU/MPS), move to CPU first
+            audio = wav.detach().cpu().numpy().squeeze()
+            return audio, 48000
+
+        return await asyncio.to_thread(_generate_sync)

backend/backends/mlx_backend.py

@@ -379,9 +379,17 @@ class MLXTTSBackend:
         return audio, sample_rate
 
 
+WHISPER_HF_REPOS = {
+    "base": "openai/whisper-base",
+    "small": "openai/whisper-small",
+    "medium": "openai/whisper-medium",
+    "large": "openai/whisper-large-v3",
+}
+
+
 class MLXSTTBackend:
     """MLX-based STT backend using mlx-audio Whisper."""
-    
+
     def __init__(self, model_size: str = "base"):
         self.model = None
         self.model_size = model_size
@@ -402,8 +410,8 @@ class MLXSTTBackend:
         """
         try:
             from huggingface_hub import constants as hf_constants
-            model_name = f"openai/whisper-{model_size}"
-            repo_cache = Path(hf_constants.HF_HUB_CACHE) / ("models--" + model_name.replace("/", "--"))
+            hf_repo = WHISPER_HF_REPOS.get(model_size, f"openai/whisper-{model_size}")
+            repo_cache = Path(hf_constants.HF_HUB_CACHE) / ("models--" + hf_repo.replace("/", "--"))
             
             if not repo_cache.exists():
                 return False
@@ -474,7 +482,7 @@ class MLXSTTBackend:
             from mlx_audio.stt import load
 
             # MLX Whisper uses the standard OpenAI models
-            model_name = f"openai/whisper-{model_size}"
+            model_name = WHISPER_HF_REPOS.get(model_size, f"openai/whisper-{model_size}")
 
             print(f"Loading MLX Whisper model {model_size}...")
 

backend/backends/pytorch_backend.py

@@ -29,9 +29,23 @@ class PyTorchTTSBackend:
         """Get the best available device."""
         if torch.cuda.is_available():
             return "cuda"
-        elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
-            # MPS can have issues, use CPU for stability
-            return "cpu"
+        # Intel Arc / Intel Xe GPU via intel-extension-for-pytorch (IPEX)
+        try:
+            import intel_extension_for_pytorch  # noqa: F401
+            if hasattr(torch, 'xpu') and torch.xpu.is_available():
+                return "xpu"
+        except ImportError:
+            pass
+        # Any GPU on Windows via DirectML (torch-directml)
+        try:
+            import torch_directml
+            if torch_directml.device_count() > 0:
+                return torch_directml.device(0)
+        except ImportError:
+            pass
+        # MPS (Apple Silicon) — kept for completeness but MLX backend is preferred
+        if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            return "cpu"  # MPS disabled for stability; MLX backend handles Apple Silicon
         return "cpu"
     
     def is_loaded(self) -> bool:
@@ -166,11 +180,21 @@ class PyTorchTTSBackend:
 
             # Load the model (tqdm is patched, but filters out non-download progress)
             try:
-                self.model = Qwen3TTSModel.from_pretrained(
-                    model_path,
-                    device_map=self.device,
-                    torch_dtype=torch.float32 if self.device == "cpu" else torch.bfloat16,
-                )
+                # Don't pass device_map on CPU: accelerate's meta-tensor mechanism
+                # causes "Cannot copy out of meta tensor" when moving to CPU.
+                # Instead load directly then call .to(device) if needed.
+                if self.device == "cpu":
+                    self.model = Qwen3TTSModel.from_pretrained(
+                        model_path,
+                        torch_dtype=torch.float32,
+                        low_cpu_mem_usage=False,
+                    )
+                else:
+                    self.model = Qwen3TTSModel.from_pretrained(
+                        model_path,
+                        device_map=self.device,
+                        torch_dtype=torch.bfloat16,
+                    )
             finally:
                 # Exit the patch context
                 tracker_context.__exit__(None, None, None)
@@ -345,9 +369,17 @@ class PyTorchTTSBackend:
         return audio, sample_rate
 
 
+WHISPER_HF_REPOS = {
+    "base": "openai/whisper-base",
+    "small": "openai/whisper-small",
+    "medium": "openai/whisper-medium",
+    "large": "openai/whisper-large-v3",
+}
+
+
 class PyTorchSTTBackend:
     """PyTorch-based STT backend using Whisper."""
-    
+
     def __init__(self, model_size: str = "base"):
         self.model = None
         self.processor = None
@@ -358,9 +390,22 @@ class PyTorchSTTBackend:
         """Get the best available device."""
         if torch.cuda.is_available():
             return "cuda"
-        elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
-            # MPS support for Whisper
-            return "cpu"  # Use CPU for stability
+        # Intel Arc / Intel Xe GPU via intel-extension-for-pytorch (IPEX)
+        try:
+            import intel_extension_for_pytorch  # noqa: F401
+            if hasattr(torch, 'xpu') and torch.xpu.is_available():
+                return "xpu"
+        except ImportError:
+            pass
+        # Any GPU on Windows via DirectML (torch-directml)
+        try:
+            import torch_directml
+            if torch_directml.device_count() > 0:
+                return torch_directml.device(0)
+        except ImportError:
+            pass
+        if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            return "cpu"  # MPS disabled for stability
         return "cpu"
     
     def is_loaded(self) -> bool:
@@ -379,18 +424,18 @@ class PyTorchSTTBackend:
         """
         try:
             from huggingface_hub import constants as hf_constants
-            model_name = f"openai/whisper-{model_size}"
-            repo_cache = Path(hf_constants.HF_HUB_CACHE) / ("models--" + model_name.replace("/", "--"))
-            
+            hf_repo = WHISPER_HF_REPOS.get(model_size, f"openai/whisper-{model_size}")
+            repo_cache = Path(hf_constants.HF_HUB_CACHE) / ("models--" + hf_repo.replace("/", "--"))
+
             if not repo_cache.exists():
                 return False
-            
+
             # Check for .incomplete files - if any exist, download is still in progress
             blobs_dir = repo_cache / "blobs"
             if blobs_dir.exists() and any(blobs_dir.glob("*.incomplete")):
                 print(f"[_is_model_cached] Found .incomplete files for whisper-{model_size}, treating as not cached")
                 return False
-            
+
             # Check that actual model weight files exist in snapshots
             snapshots_dir = repo_cache / "snapshots"
             if snapshots_dir.exists():
@@ -401,12 +446,12 @@ class PyTorchSTTBackend:
                 if not has_weights:
                     print(f"[_is_model_cached] No model weights found for whisper-{model_size}, treating as not cached")
                     return False
-            
+
             return True
         except Exception as e:
             print(f"[_is_model_cached] Error checking cache for whisper-{model_size}: {e}")
             return False
-    
+
     async def load_model_async(self, model_size: Optional[str] = None):
         """
         Lazy load the Whisper model.
@@ -457,7 +502,7 @@ class PyTorchSTTBackend:
             # Import transformers
             from transformers import WhisperProcessor, WhisperForConditionalGeneration
 
-            model_name = f"openai/whisper-{model_size}"
+            model_name = WHISPER_HF_REPOS.get(model_size, f"openai/whisper-{model_size}")
             print(f"[DEBUG] Model name: {model_name}")
 
             print(f"Loading Whisper model {model_size} on {self.device}...")

backend/build_binary.py

@@ -1,8 +1,13 @@
 """
 PyInstaller build script for creating standalone Python server binary.
+
+Usage:
+    python build_binary.py           # Build default (CPU) server binary
+    python build_binary.py --cuda    # Build CUDA-enabled server binary
 """
 
 import PyInstaller.__main__
+import argparse
 import os
 import platform
 from pathlib import Path
@@ -13,15 +18,22 @@ def is_apple_silicon():
     return platform.system() == "Darwin" and platform.machine() == "arm64"
 
 
-def build_server():
-    """Build Python server as standalone binary."""
+def build_server(cuda=False):
+    """Build Python server as standalone binary.
+
+    Args:
+        cuda: If True, build with CUDA support and name the binary
+              voicebox-server-cuda instead of voicebox-server.
+    """
     backend_dir = Path(__file__).parent
 
+    binary_name = 'voicebox-server-cuda' if cuda else 'voicebox-server'
+
     # PyInstaller arguments
     args = [
         'server.py',  # Use server.py as entry point instead of main.py
         '--onefile',
-        '--name', 'voicebox-server',
+        '--name', binary_name,
     ]
 
     # Add local qwen_tts path if specified (for editable installs)
@@ -49,6 +61,7 @@ def build_server():
         '--hidden-import', 'backend.utils.progress',
         '--hidden-import', 'backend.utils.hf_progress',
         '--hidden-import', 'backend.utils.validation',
+        '--hidden-import', 'backend.cuda_download',
         '--hidden-import', 'torch',
         '--hidden-import', 'transformers',
         '--hidden-import', 'fastapi',
@@ -70,8 +83,16 @@ def build_server():
         '--collect-submodules', 'jaraco',
     ])
 
-    # Add MLX-specific imports if building on Apple Silicon
-    if is_apple_silicon():
+    # Add CUDA-specific hidden imports
+    if cuda:
+        print("Building with CUDA support")
+        args.extend([
+            '--hidden-import', 'torch.cuda',
+            '--hidden-import', 'torch.backends.cudnn',
+        ])
+
+    # Add MLX-specific imports if building on Apple Silicon (never for CUDA builds)
+    if is_apple_silicon() and not cuda:
         print("Building for Apple Silicon - including MLX dependencies")
         args.extend([
             '--hidden-import', 'backend.backends.mlx_backend',
@@ -83,11 +104,15 @@ def build_server():
             '--hidden-import', 'mlx_audio.stt',
             '--collect-submodules', 'mlx',
             '--collect-submodules', 'mlx_audio',
-            # Collect MLX data files including Metal shader libraries (.metallib)
-            '--collect-data', 'mlx',
-            '--collect-data', 'mlx_audio',
+            # Use --collect-all so PyInstaller bundles both data files AND
+            # native shared libraries (.dylib, .metallib) for MLX.
+            # Previously only --collect-data was used, which caused MLX to
+            # raise OSError at runtime inside the bundled binary because
+            # the Metal shader libraries were missing.
+            '--collect-all', 'mlx',
+            '--collect-all', 'mlx_audio',
         ])
-    else:
+    elif not cuda:
         print("Building for non-Apple Silicon platform - PyTorch only")
 
     args.extend([
@@ -101,8 +126,15 @@ def build_server():
     # Run PyInstaller
     PyInstaller.__main__.run(args)
     
-    print(f"Binary built in {backend_dir / 'dist' / 'voicebox-server'}")
+    print(f"Binary built in {backend_dir / 'dist' / binary_name}")
 
 
 if __name__ == '__main__':
-    build_server()
+    parser = argparse.ArgumentParser(description="Build voicebox-server binary")
+    parser.add_argument(
+        '--cuda',
+        action='store_true',
+        help="Build CUDA-enabled binary (voicebox-server-cuda)",
+    )
+    cli_args = parser.parse_args()
+    build_server(cuda=cli_args.cuda)

backend/config.py

@@ -4,8 +4,17 @@ Configuration module for voicebox backend.
 Handles data directory configuration for production bundling.
 """
 
+import os
 from pathlib import Path
 
+# Allow users to override the HuggingFace model download directory.
+# Set VOICEBOX_MODELS_DIR to an absolute path before starting the server.
+# This sets HF_HUB_CACHE so all huggingface_hub downloads go to that path.
+_custom_models_dir = os.environ.get("VOICEBOX_MODELS_DIR")
+if _custom_models_dir:
+    os.environ["HF_HUB_CACHE"] = _custom_models_dir
+    print(f"[config] Model download path set to: {_custom_models_dir}")
+
 # Default data directory (used in development)
 _data_dir = Path("data")
 

backend/cuda_download.py

@@ -0,0 +1,198 @@
+"""
+CUDA backend binary download, assembly, and verification.
+
+Downloads split parts of the CUDA-enabled voicebox-server binary from
+GitHub Releases, reassembles them, verifies integrity via SHA-256,
+and places the binary in the app's data directory for use on next
+backend restart.
+"""
+
+import hashlib
+import logging
+import os
+import sys
+from pathlib import Path
+from typing import Optional
+
+from .config import get_data_dir
+from .utils.progress import get_progress_manager
+from . import __version__
+
+logger = logging.getLogger(__name__)
+
+GITHUB_RELEASES_URL = "https://github.com/jamiepine/voicebox/releases/download"
+
+PROGRESS_KEY = "cuda-backend"
+
+
+def get_backends_dir() -> Path:
+    """Directory where downloaded backend binaries are stored."""
+    d = get_data_dir() / "backends"
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
+def get_cuda_binary_name() -> str:
+    """Platform-specific CUDA binary filename."""
+    if sys.platform == "win32":
+        return "voicebox-server-cuda.exe"
+    return "voicebox-server-cuda"
+
+
+def get_cuda_binary_path() -> Optional[Path]:
+    """Return path to CUDA binary if it exists."""
+    p = get_backends_dir() / get_cuda_binary_name()
+    if p.exists():
+        return p
+    return None
+
+
+def is_cuda_active() -> bool:
+    """Check if the current process is the CUDA binary.
+
+    The CUDA binary sets this env var on startup (see server.py).
+    """
+    return os.environ.get("VOICEBOX_BACKEND_VARIANT") == "cuda"
+
+
+def get_cuda_status() -> dict:
+    """Get current CUDA backend status for the API."""
+    progress_manager = get_progress_manager()
+    cuda_path = get_cuda_binary_path()
+    progress = progress_manager.get_progress(PROGRESS_KEY)
+
+    return {
+        "available": cuda_path is not None,
+        "active": is_cuda_active(),
+        "binary_path": str(cuda_path) if cuda_path else None,
+        "downloading": progress is not None and progress.get("status") == "downloading",
+        "download_progress": progress,
+    }
+
+
+async def download_cuda_binary(version: Optional[str] = None):
+    """Download the CUDA backend binary from GitHub Releases.
+
+    Downloads split parts listed in a manifest file, concatenates them,
+    and verifies the SHA-256 checksum for integrity. Atomic write
+    (temp file -> rename).
+
+    Args:
+        version: Version tag (e.g. "v0.2.0"). Defaults to current app version.
+    """
+    import httpx
+
+    if version is None:
+        version = f"v{__version__}"
+
+    progress = get_progress_manager()
+    binary_name = get_cuda_binary_name()
+    dest_dir = get_backends_dir()
+    final_path = dest_dir / binary_name
+    temp_path = dest_dir / f"{binary_name}.download"
+
+    # Clean up any leftover partial download
+    if temp_path.exists():
+        temp_path.unlink()
+
+    logger.info(f"Starting CUDA backend download for {version}")
+    progress.update_progress(
+        PROGRESS_KEY, current=0, total=0,
+        filename="Fetching manifest...", status="downloading",
+    )
+
+    base_url = f"{GITHUB_RELEASES_URL}/{version}"
+    stem = Path(binary_name).stem  # voicebox-server-cuda
+
+    try:
+        async with httpx.AsyncClient(follow_redirects=True, timeout=30.0) as client:
+            # Fetch the manifest (list of split part filenames)
+            manifest_url = f"{base_url}/{stem}.manifest"
+            manifest_resp = await client.get(manifest_url)
+            manifest_resp.raise_for_status()
+            parts = [p.strip() for p in manifest_resp.text.strip().splitlines() if p.strip()]
+
+            if not parts:
+                raise ValueError("Empty manifest — no split parts found")
+
+            logger.info(f"Found {len(parts)} split parts to download")
+
+            # Fetch expected checksum (optional — for integrity verification)
+            expected_sha = None
+            try:
+                sha_url = f"{base_url}/{stem}.sha256"
+                sha_resp = await client.get(sha_url)
+                if sha_resp.status_code == 200:
+                    # Format: "sha256hex  filename\n"
+                    expected_sha = sha_resp.text.strip().split()[0]
+                    logger.info(f"Expected SHA-256: {expected_sha[:16]}...")
+            except Exception as e:
+                logger.warning(f"Could not fetch checksum file — skipping verification: {e}")
+
+            # Download and concatenate parts
+            total_downloaded = 0
+            with open(temp_path, "wb") as f:
+                for i, part_name in enumerate(parts):
+                    part_url = f"{base_url}/{part_name}"
+                    logger.info(f"Downloading part {i + 1}/{len(parts)}: {part_name}")
+
+                    async with client.stream("GET", part_url) as response:
+                        response.raise_for_status()
+                        async for chunk in response.aiter_bytes(chunk_size=1024 * 1024):
+                            f.write(chunk)
+                            total_downloaded += len(chunk)
+                            progress.update_progress(
+                                PROGRESS_KEY, current=total_downloaded, total=0,
+                                filename=f"Part {i + 1}/{len(parts)}",
+                                status="downloading",
+                            )
+
+        # Verify integrity if checksum was available
+        if expected_sha:
+            progress.update_progress(
+                PROGRESS_KEY, current=total_downloaded, total=total_downloaded,
+                filename="Verifying integrity...", status="downloading",
+            )
+            sha256 = hashlib.sha256()
+            with open(temp_path, "rb") as f:
+                while True:
+                    chunk = f.read(1024 * 1024)
+                    if not chunk:
+                        break
+                    sha256.update(chunk)
+
+            actual = sha256.hexdigest()
+            if actual != expected_sha:
+                raise ValueError(
+                    f"Integrity check failed: expected {expected_sha[:16]}..., "
+                    f"got {actual[:16]}..."
+                )
+            logger.info(f"Integrity verified: {actual[:16]}...")
+
+        # Atomic move into place (replace handles existing target on all platforms)
+        temp_path.replace(final_path)
+
+        # Make executable on Unix
+        if sys.platform != "win32":
+            final_path.chmod(0o755)
+
+        logger.info(f"CUDA backend downloaded to {final_path}")
+        progress.mark_complete(PROGRESS_KEY)
+
+    except Exception as e:
+        # Clean up on failure
+        if temp_path.exists():
+            temp_path.unlink()
+        logger.error(f"CUDA backend download failed: {e}")
+        progress.mark_error(PROGRESS_KEY, str(e))
+        raise
+
+
+async def delete_cuda_binary() -> bool:
+    """Delete the downloaded CUDA binary. Returns True if deleted."""
+    path = get_cuda_binary_path()
+    if path and path.exists():
+        path.unlink()
+        logger.info(f"Deleted CUDA binary: {path}")
+        return True
+    return False

backend/main.py

@@ -22,6 +22,24 @@ import uuid
 import asyncio
 import signal
 import os
+from urllib.parse import quote
+
+
+def _safe_content_disposition(disposition_type: str, filename: str) -> str:
+    """Build a Content-Disposition header that is safe for non-ASCII filenames.
+
+    Uses RFC 5987 ``filename*`` parameter so that browsers can decode
+    UTF-8 filenames while the ``filename`` fallback stays ASCII-only.
+    """
+    ascii_name = "".join(
+        c for c in filename if c.isascii() and (c.isalnum() or c in " -_.")
+    ).strip() or "download"
+    utf8_name = quote(filename, safe="")
+    return (
+        f'{disposition_type}; filename="{ascii_name}"; '
+        f"filename*=UTF-8''{utf8_name}"
+    )
+
 
 from . import database, models, profiles, history, tts, transcribe, config, export_import, channels, stories, __version__
 from .database import get_db, Generation as DBGeneration, VoiceProfile as DBVoiceProfile
@@ -30,6 +48,18 @@ from .utils.tasks import get_task_manager
 from .utils.cache import clear_voice_prompt_cache
 from .platform_detect import get_backend_type
 
+# Keep references to fire-and-forget background tasks to prevent GC
+_background_tasks: set = set()
+
+
+def _create_background_task(coro) -> asyncio.Task:
+    """Create a background task and prevent it from being garbage collected."""
+    task = asyncio.create_task(coro)
+    _background_tasks.add(task)
+    task.add_done_callback(_background_tasks.discard)
+    return task
+
+
 app = FastAPI(
     title="voicebox API",
     description="Production-quality Qwen3-TTS voice cloning API",
@@ -77,10 +107,39 @@ async def health():
     tts_model = tts.get_tts_model()
     backend_type = get_backend_type()
 
-    # Check for GPU availability (CUDA or MPS)
+    # Check for GPU availability (CUDA, MPS, Intel Arc XPU, or DirectML)
     has_cuda = torch.cuda.is_available()
     has_mps = hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()
-    gpu_available = has_cuda or has_mps
+
+    # Intel Arc / Intel Xe via intel-extension-for-pytorch (IPEX)
+    has_xpu = False
+    xpu_name = None
+    try:
+        import intel_extension_for_pytorch as ipex  # noqa: F401
+        if hasattr(torch, 'xpu') and torch.xpu.is_available():
+            has_xpu = True
+            try:
+                xpu_name = torch.xpu.get_device_name(0)
+            except Exception:
+                xpu_name = "Intel GPU"
+    except ImportError:
+        pass
+
+    # DirectML backend (torch-directml) for any Windows GPU
+    has_directml = False
+    directml_name = None
+    try:
+        import torch_directml
+        if torch_directml.device_count() > 0:
+            has_directml = True
+            try:
+                directml_name = torch_directml.device_name(0)
+            except Exception:
+                directml_name = "DirectML GPU"
+    except ImportError:
+        pass
+
+    gpu_available = has_cuda or has_mps or has_xpu or has_directml or backend_type == "mlx"
 
     gpu_type = None
     if has_cuda:
@@ -89,6 +148,10 @@ async def health():
         gpu_type = "MPS (Apple Silicon)"
     elif backend_type == "mlx":
         gpu_type = "Metal (Apple Silicon via MLX)"
+    elif has_xpu:
+        gpu_type = f"XPU ({xpu_name})"
+    elif has_directml:
+        gpu_type = f"DirectML ({directml_name})"
 
     vram_used = None
     if has_cuda:
@@ -155,6 +218,7 @@ async def health():
         gpu_type=gpu_type,
         vram_used_mb=vram_used,
         backend_type=backend_type,
+        backend_variant=os.environ.get("VOICEBOX_BACKEND_VARIANT", "cpu"),
     )
 
 
@@ -252,12 +316,17 @@ async def add_profile_sample(
     db: Session = Depends(get_db),
 ):
     """Add a sample to a voice profile."""
-    # Save uploaded file to temporary location
-    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
+    # Preserve the uploaded file's extension so librosa can detect format correctly.
+    # Defaulting to .wav was causing soundfile to reject MP3/WebM content as invalid WAV.
+    _allowed_audio_exts = {'.wav', '.mp3', '.m4a', '.ogg', '.flac', '.aac', '.webm', '.opus'}
+    _uploaded_ext = Path(file.filename or '').suffix.lower()
+    file_suffix = _uploaded_ext if _uploaded_ext in _allowed_audio_exts else '.wav'
+
+    with tempfile.NamedTemporaryFile(suffix=file_suffix, delete=False) as tmp:
         content = await file.read()
         tmp.write(content)
         tmp_path = tmp.name
-    
+
     try:
         sample = await profiles.add_profile_sample(
             profile_id,
@@ -268,6 +337,8 @@ async def add_profile_sample(
         return sample
     except ValueError as e:
         raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Failed to process audio file: {str(e)}")
     finally:
         # Clean up temp file
         Path(tmp_path).unlink(missing_ok=True)
@@ -388,7 +459,7 @@ async def export_profile(
             io.BytesIO(zip_bytes),
             media_type="application/zip",
             headers={
-                "Content-Disposition": f'attachment; filename="{filename}"'
+                "Content-Disposition": _safe_content_disposition("attachment", filename)
             }
         )
     except ValueError as e:
@@ -542,25 +613,18 @@ async def generate_speech(
         if not profile:
             raise HTTPException(status_code=404, detail="Profile not found")
         
-        # Create voice prompt from profile
-        voice_prompt = await profiles.create_voice_prompt_for_profile(
-            data.profile_id,
-            db,
-        )
-        
         # Generate audio
-        tts_model = tts.get_tts_model()
-        # Load the requested model size if different from current (async to not block)
+        from .backends import get_tts_backend_for_engine
+
+        engine = data.engine or "qwen"
+        tts_model = get_tts_backend_for_engine(engine)
+
+        # Resolve model size (only relevant for Qwen engine)
         model_size = data.model_size or "1.7B"
 
         # Check if model needs to be downloaded first
-        model_path = tts_model._get_model_path(model_size)
-        if model_path.startswith("Qwen/"):
-            # Model not cached - check if it exists remotely or needs download
-            from huggingface_hub import constants as hf_constants
-            repo_cache = Path(hf_constants.HF_HUB_CACHE) / ("models--" + model_path.replace("/", "--"))
-            if not repo_cache.exists():
-                # Start download in background
+        if engine == "qwen":
+            if not tts_model._is_model_cached(model_size):
                 model_name = f"qwen-tts-{model_size}"
 
                 async def download_model_background():
@@ -570,19 +634,51 @@ async def generate_speech(
                         task_manager.error_download(model_name, str(e))
 
                 task_manager.start_download(model_name)
-                asyncio.create_task(download_model_background())
+                _create_background_task(download_model_background())
 
-                # Return 202 Accepted with download info
                 raise HTTPException(
                     status_code=202,
                     detail={
                         "message": f"Model {model_size} is being downloaded. Please wait and try again.",
                         "model_name": model_name,
-                        "downloading": True
-                    }
+                        "downloading": True,
+                    },
                 )
 
-        await tts_model.load_model_async(model_size)
+            # Load (or switch to) the requested model
+            await tts_model.load_model_async(model_size)
+        elif engine == "luxtts":
+            if not tts_model._is_model_cached():
+                model_name = "luxtts"
+
+                async def download_luxtts_background():
+                    try:
+                        await tts_model.load_model()
+                    except Exception as e:
+                        task_manager.error_download(model_name, str(e))
+
+                task_manager.start_download(model_name)
+                _create_background_task(download_luxtts_background())
+
+                raise HTTPException(
+                    status_code=202,
+                    detail={
+                        "message": "LuxTTS model is being downloaded. Please wait and try again.",
+                        "model_name": model_name,
+                        "downloading": True,
+                    },
+                )
+
+            await tts_model.load_model()
+
+        # Create voice prompt from profile
+        voice_prompt = await profiles.create_voice_prompt_for_profile(
+            data.profile_id,
+            db,
+            use_cache=True,
+            engine=engine,
+        )
+
         audio, sample_rate = await tts_model.generate(
             data.text,
             voice_prompt,
@@ -625,6 +721,70 @@ async def generate_speech(
         raise HTTPException(status_code=500, detail=str(e))
 
 
+@app.post("/generate/stream")
+async def stream_speech(
+    data: models.GenerationRequest,
+    db: Session = Depends(get_db),
+):
+    """
+    Generate speech and stream the WAV audio directly without saving to disk.
+
+    Returns raw WAV bytes via a StreamingResponse so the client can start
+    playing audio before the entire file has been received.  This endpoint
+    does NOT create a history entry — use /generate for that.
+    """
+    from .backends import get_tts_backend_for_engine
+
+    profile = await profiles.get_profile(data.profile_id, db)
+    if not profile:
+        raise HTTPException(status_code=404, detail="Profile not found")
+
+    engine = data.engine or "qwen"
+    tts_model = get_tts_backend_for_engine(engine)
+    model_size = data.model_size or "1.7B"
+
+    if engine == "qwen":
+        if not tts_model._is_model_cached(model_size):
+            raise HTTPException(
+                status_code=400,
+                detail=f"Model {model_size} is not downloaded yet. Use /generate to trigger a download.",
+            )
+        await tts_model.load_model_async(model_size)
+    elif engine == "luxtts":
+        if not tts_model._is_model_cached():
+            raise HTTPException(
+                status_code=400,
+                detail="LuxTTS model is not downloaded yet. Use /generate to trigger a download.",
+            )
+        await tts_model.load_model()
+
+    voice_prompt = await profiles.create_voice_prompt_for_profile(
+        data.profile_id, db, engine=engine,
+    )
+
+    audio, sample_rate = await tts_model.generate(
+        data.text,
+        voice_prompt,
+        data.language,
+        data.seed,
+        data.instruct,
+    )
+
+    wav_bytes = tts.audio_to_wav_bytes(audio, sample_rate)
+
+    async def _wav_stream():
+        # Yield in chunks so large responses don't block the event loop
+        chunk_size = 64 * 1024  # 64 KB
+        for i in range(0, len(wav_bytes), chunk_size):
+            yield wav_bytes[i : i + chunk_size]
+
+    return StreamingResponse(
+        _wav_stream(),
+        media_type="audio/wav",
+        headers={"Content-Disposition": 'attachment; filename="speech.wav"'},
+    )
+
+
 # ============================================
 # HISTORY ENDPOINTS
 # ============================================
@@ -753,7 +913,7 @@ async def export_generation(
             io.BytesIO(zip_bytes),
             media_type="application/zip",
             headers={
-                "Content-Disposition": f'attachment; filename="{filename}"'
+                "Content-Disposition": _safe_content_disposition("attachment", filename)
             }
         )
     except ValueError as e:
@@ -786,7 +946,7 @@ async def export_generation_audio(
         audio_path,
         media_type="audio/wav",
         headers={
-            "Content-Disposition": f'attachment; filename="{filename}"'
+            "Content-Disposition": _safe_content_disposition("attachment", filename)
         }
     )
 
@@ -818,7 +978,11 @@ async def transcribe_audio(
 
         # Check if Whisper model is downloaded (uses default size "base")
         model_size = whisper_model.model_size
-        model_name = f"openai/whisper-{model_size}"
+        # Map model sizes to HF repo IDs (whisper-large needs -v3 suffix)
+        whisper_hf_repos = {
+            "large": "openai/whisper-large-v3",
+        }
+        model_name = whisper_hf_repos.get(model_size, f"openai/whisper-{model_size}")
 
         # Check if model is cached
         from huggingface_hub import constants as hf_constants
@@ -834,7 +998,7 @@ async def transcribe_audio(
                     get_task_manager().error_download(progress_model_name, str(e))
 
             get_task_manager().start_download(progress_model_name)
-            asyncio.create_task(download_whisper_background())
+            _create_background_task(download_whisper_background())
 
             # Return 202 Accepted
             raise HTTPException(
@@ -1054,7 +1218,7 @@ async def export_story_audio(
             io.BytesIO(audio_bytes),
             media_type="audio/wav",
             headers={
-                "Content-Disposition": f'attachment; filename="{filename}"'
+                "Content-Disposition": _safe_content_disposition("attachment", filename)
             }
         )
     except HTTPException:
@@ -1196,15 +1360,24 @@ async def get_model_status():
         whisper_base_id = "openai/whisper-base"
         whisper_small_id = "openai/whisper-small"
         whisper_medium_id = "openai/whisper-medium"
-        whisper_large_id = "openai/whisper-large"
+        whisper_large_id = "openai/whisper-large-v3"
     else:
         tts_1_7b_id = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
         tts_0_6b_id = "Qwen/Qwen3-TTS-12Hz-0.6B-Base"
         whisper_base_id = "openai/whisper-base"
         whisper_small_id = "openai/whisper-small"
         whisper_medium_id = "openai/whisper-medium"
-        whisper_large_id = "openai/whisper-large"
+        whisper_large_id = "openai/whisper-large-v3"
     
+    # Check if LuxTTS backend is loaded
+    def check_luxtts_loaded():
+        try:
+            from .backends import get_tts_backend_for_engine
+            backend = get_tts_backend_for_engine("luxtts")
+            return backend.is_loaded()
+        except Exception:
+            return False
+
     model_configs = [
         {
             "model_name": "qwen-tts-1.7B",
@@ -1220,6 +1393,13 @@ async def get_model_status():
             "model_size": "0.6B",
             "check_loaded": lambda: check_tts_loaded("0.6B"),
         },
+        {
+            "model_name": "luxtts",
+            "display_name": "LuxTTS (Fast, CPU-friendly)",
+            "hf_repo_id": "YatharthS/LuxTTS",
+            "model_size": "default",
+            "check_loaded": check_luxtts_loaded,
+        },
         {
             "model_name": "whisper-base",
             "display_name": "Whisper Base",
@@ -1402,6 +1582,7 @@ async def get_model_status():
 async def trigger_model_download(request: models.ModelDownloadRequest):
     """Trigger download of a specific model."""
     import asyncio
+    from .backends import get_tts_backend_for_engine
     
     task_manager = get_task_manager()
     progress_manager = get_progress_manager()
@@ -1415,6 +1596,10 @@ async def trigger_model_download(request: models.ModelDownloadRequest):
             "model_size": "0.6B",
             "load_func": lambda: tts.get_tts_model().load_model("0.6B"),
         },
+        "luxtts": {
+            "model_size": "default",
+            "load_func": lambda: get_tts_backend_for_engine("luxtts").load_model(),
+        },
         "whisper-base": {
             "model_size": "base",
             "load_func": lambda: transcribe.get_whisper_model().load_model("base"),
@@ -1466,12 +1651,48 @@ async def trigger_model_download(request: models.ModelDownloadRequest):
     )
 
     # Start download in background task (don't await)
-    asyncio.create_task(download_in_background())
+    _create_background_task(download_in_background())
 
     # Return immediately - frontend should poll progress endpoint
     return {"message": f"Model {request.model_name} download started"}
 
 
+@app.post("/models/download/cancel")
+async def cancel_model_download(request: models.ModelDownloadRequest):
+    """Cancel or dismiss an errored/stale download task."""
+    task_manager = get_task_manager()
+    progress_manager = get_progress_manager()
+
+    removed = task_manager.cancel_download(request.model_name)
+
+    # Also clear progress state so the model doesn't show as downloading
+    progress_removed = False
+    with progress_manager._lock:
+        if request.model_name in progress_manager._progress:
+            del progress_manager._progress[request.model_name]
+            progress_removed = True
+
+    if removed or progress_removed:
+        return {"message": f"Download task for {request.model_name} cancelled"}
+    return {"message": f"No active task found for {request.model_name}"}
+
+
+@app.post("/tasks/clear")
+async def clear_all_tasks():
+    """Clear all download tasks and progress state. Does not delete downloaded files."""
+    task_manager = get_task_manager()
+    progress_manager = get_progress_manager()
+
+    task_manager.clear_all()
+
+    with progress_manager._lock:
+        progress_manager._progress.clear()
+        progress_manager._last_notify_time.clear()
+        progress_manager._last_notify_progress.clear()
+
+    return {"message": "All task state cleared"}
+
+
 @app.delete("/models/{model_name}")
 async def delete_model(model_name: str):
     """Delete a downloaded model from the HuggingFace cache."""
@@ -1491,6 +1712,11 @@ async def delete_model(model_name: str):
             "model_size": "0.6B",
             "model_type": "tts",
         },
+        "luxtts": {
+            "hf_repo_id": "YatharthS/LuxTTS",
+            "model_size": "default",
+            "model_type": "luxtts",
+        },
         "whisper-base": {
             "hf_repo_id": "openai/whisper-base",
             "model_size": "base",
@@ -1507,12 +1733,12 @@ async def delete_model(model_name: str):
             "model_type": "whisper",
         },
         "whisper-large": {
-            "hf_repo_id": "openai/whisper-large",
+            "hf_repo_id": "openai/whisper-large-v3",
             "model_size": "large",
             "model_type": "whisper",
         },
     }
-    
+
     if model_name not in model_configs:
         raise HTTPException(status_code=400, detail=f"Unknown model: {model_name}")
     
@@ -1525,6 +1751,11 @@ async def delete_model(model_name: str):
             tts_model = tts.get_tts_model()
             if tts_model.is_loaded() and tts_model.model_size == config["model_size"]:
                 tts.unload_tts_model()
+        elif config["model_type"] == "luxtts":
+            from .backends import get_tts_backend_for_engine
+            luxtts = get_tts_backend_for_engine("luxtts")
+            if luxtts.is_loaded():
+                luxtts.unload_model()
         elif config["model_type"] == "whisper":
             whisper_model = transcribe.get_whisper_model()
             if whisper_model.is_loaded() and whisper_model.model_size == config["model_size"]:
@@ -1596,10 +1827,18 @@ async def get_active_tasks():
         progress = progress_map.get(model_name)
         
         if task:
+            # Prefer task error, fall back to progress manager error
+            error = task.error
+            if not error:
+                with progress_manager._lock:
+                    pm_data = progress_manager._progress.get(model_name)
+                    if pm_data:
+                        error = pm_data.get("error")
             active_downloads.append(models.ActiveDownloadTask(
                 model_name=model_name,
                 status=task.status,
                 started_at=task.started_at,
+                error=error,
             ))
         elif progress:
             # Progress exists but no task - create from progress data
@@ -1616,6 +1855,7 @@ async def get_active_tasks():
                 model_name=model_name,
                 status=progress.get("status", "downloading"),
                 started_at=started_at,
+                error=progress.get("error"),
             ))
     
     # Get active generations
@@ -1634,6 +1874,75 @@ async def get_active_tasks():
     )
 
 
+# ============================================
+# CUDA BACKEND MANAGEMENT
+# ============================================
+
+@app.get("/backend/cuda-status")
+async def get_cuda_status():
+    """Get CUDA backend download/availability status."""
+    from . import cuda_download
+    return cuda_download.get_cuda_status()
+
+
+@app.post("/backend/download-cuda")
+async def download_cuda_backend():
+    """Download the CUDA backend binary. Returns immediately; track progress via SSE."""
+    from . import cuda_download
+
+    # Check if already downloaded
+    if cuda_download.get_cuda_binary_path() is not None:
+        raise HTTPException(status_code=409, detail="CUDA backend already downloaded")
+
+    async def _download():
+        try:
+            await cuda_download.download_cuda_binary()
+        except Exception as e:
+            import logging
+            logging.getLogger(__name__).error(f"CUDA download failed: {e}")
+
+    _create_background_task(_download())
+    return {"message": "CUDA backend download started", "progress_key": "cuda-backend"}
+
+
+@app.delete("/backend/cuda")
+async def delete_cuda_backend():
+    """Delete the downloaded CUDA backend binary."""
+    from . import cuda_download
+
+    if cuda_download.is_cuda_active():
+        raise HTTPException(
+            status_code=409,
+            detail="Cannot delete CUDA backend while it is active. Switch to CPU first.",
+        )
+
+    deleted = await cuda_download.delete_cuda_binary()
+    if not deleted:
+        raise HTTPException(status_code=404, detail="No CUDA backend found to delete")
+
+    return {"message": "CUDA backend deleted"}
+
+
+@app.get("/backend/cuda-progress")
+async def get_cuda_download_progress():
+    """Get CUDA backend download progress via Server-Sent Events."""
+    progress_manager = get_progress_manager()
+
+    async def event_generator():
+        async for event in progress_manager.subscribe("cuda-backend"):
+            yield event
+
+    return StreamingResponse(
+        event_generator(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache",
+            "Connection": "keep-alive",
+            "X-Accel-Buffering": "no",
+        },
+    )
+
+
 # ============================================
 # STARTUP & SHUTDOWN
 # ============================================

backend/models.py

@@ -57,6 +57,7 @@ class GenerationRequest(BaseModel):
     seed: Optional[int] = Field(None, ge=0)
     model_size: Optional[str] = Field(default="1.7B", pattern="^(1\\.7B|0\\.6B)$")
     instruct: Optional[str] = Field(None, max_length=500)
+    engine: Optional[str] = Field(default="qwen", pattern="^(qwen|luxtts)$")
 
 
 class GenerationResponse(BaseModel):
@@ -127,6 +128,7 @@ class HealthResponse(BaseModel):
     gpu_type: Optional[str] = None  # GPU type (CUDA, MPS, or None)
     vram_used_mb: Optional[float] = None
     backend_type: Optional[str] = None  # Backend type (mlx or pytorch)
+    backend_variant: Optional[str] = None  # Binary variant (cpu or cuda)
 
 
 class ModelStatus(BaseModel):
@@ -154,6 +156,7 @@ class ActiveDownloadTask(BaseModel):
     model_name: str
     status: str
     started_at: datetime
+    error: Optional[str] = None
 
 
 class ActiveGenerationTask(BaseModel):

backend/platform_detect.py

@@ -19,15 +19,17 @@ def is_apple_silicon() -> bool:
 def get_backend_type() -> Literal["mlx", "pytorch"]:
     """
     Detect the best backend for the current platform.
-    
+
     Returns:
-        "mlx" on Apple Silicon (if MLX is available), "pytorch" otherwise
+        "mlx" on Apple Silicon (if MLX is available and functional), "pytorch" otherwise
     """
     if is_apple_silicon():
         try:
-            import mlx
+            import mlx.core  # noqa: F401 — triggers native lib loading
             return "mlx"
-        except ImportError:
-            # MLX not installed, fallback to PyTorch
+        except (ImportError, OSError, RuntimeError):
+            # MLX not installed, or native libraries failed to load inside a
+            # PyInstaller bundle (OSError on missing .dylib / .metallib).
+            # Fall through to PyTorch.
             return "pytorch"
     return "pytorch"

backend/profiles.py

@@ -327,6 +327,7 @@ async def create_voice_prompt_for_profile(
     profile_id: str,
     db: Session,
     use_cache: bool = True,
+    engine: str = "qwen",
 ) -> dict:
     """
     Create a combined voice prompt from all samples in a profile.
@@ -335,17 +336,20 @@ async def create_voice_prompt_for_profile(
         profile_id: Profile ID
         db: Database session
         use_cache: Whether to use cached prompts
+        engine: TTS engine to create prompt for ("qwen" or "luxtts")
 
     Returns:
         Voice prompt dictionary
     """
+    from .backends import get_tts_backend_for_engine
+
     # Get all samples for profile
     samples = db.query(DBProfileSample).filter_by(profile_id=profile_id).all()
 
     if not samples:
         raise ValueError(f"No samples found for profile {profile_id}")
 
-    tts_model = get_tts_model()
+    tts_model = get_tts_backend_for_engine(engine)
 
     if len(samples) == 1:
         # Single sample - use directly

backend/requirements.txt

@@ -9,15 +9,26 @@ alembic>=1.13.0
 
 # ML models
 torch>=2.1.0
-transformers>=4.36.0
+transformers>=4.36.0,<=4.57.6
 accelerate>=0.26.0
 huggingface_hub>=0.20.0
 qwen-tts>=0.0.5
 
+# LuxTTS (voice cloning engine)
+# piper-phonemize needs custom index (no PyPI wheels)
+--find-links https://k2-fsa.github.io/icefall/piper_phonemize.html
+# linacodec is a git-only dep of Zipvoice (uv-only source, pip can't resolve it)
+linacodec @ git+https://github.com/ysharma3501/LinaCodec.git
+Zipvoice @ git+https://github.com/ysharma3501/LuxTTS.git
+
 # Audio processing
 librosa>=0.10.0
 soundfile>=0.12.0
 numpy>=1.24.0
+numba>=0.60.0,<0.61.0
+
+# HTTP client (for CUDA backend download)
+httpx>=0.27.0
 
 # Utilities
 python-multipart>=0.0.6

backend/server.py

@@ -64,7 +64,29 @@ if __name__ == "__main__":
             default=None,
             help="Data directory for database, profiles, and generated audio",
         )
+        parser.add_argument(
+            "--version",
+            action="store_true",
+            help="Print version and exit",
+        )
         args = parser.parse_args()
+
+        if args.version:
+            from backend import __version__
+            print(f"voicebox-server {__version__}")
+            sys.exit(0)
+
+        # Detect backend variant from binary name
+        # voicebox-server-cuda → sets VOICEBOX_BACKEND_VARIANT=cuda
+        import os
+        binary_name = os.path.basename(sys.executable).lower()
+        if "cuda" in binary_name:
+            os.environ["VOICEBOX_BACKEND_VARIANT"] = "cuda"
+            logger.info("Backend variant: CUDA")
+        else:
+            os.environ["VOICEBOX_BACKEND_VARIANT"] = "cpu"
+            logger.info("Backend variant: CPU")
+
         logger.info(f"Parsed arguments: host={args.host}, port={args.port}, data_dir={args.data_dir}")
 
         # Set data directory if provided

backend/tts.py

@@ -32,11 +32,3 @@ def audio_to_wav_bytes(audio: np.ndarray, sample_rate: int) -> bytes:
     sf.write(buffer, audio, sample_rate, format="WAV")
     buffer.seek(0)
     return buffer.read()
-
-
-def audio_to_wav_bytes(audio: np.ndarray, sample_rate: int) -> bytes:
-    """Convert audio array to WAV bytes."""
-    buffer = io.BytesIO()
-    sf.write(buffer, audio, sample_rate, format="WAV")
-    buffer.seek(0)
-    return buffer.read()

backend/utils/tasks.py

@@ -72,6 +72,15 @@ class TaskManager:
         """Get all active generations."""
         return list(self._active_generations.values())
     
+    def cancel_download(self, model_name: str) -> bool:
+        """Cancel/dismiss a download task (removes it from active list)."""
+        return self._active_downloads.pop(model_name, None) is not None
+
+    def clear_all(self) -> None:
+        """Clear all download and generation tasks."""
+        self._active_downloads.clear()
+        self._active_generations.clear()
+
     def is_download_active(self, model_name: str) -> bool:
         """Check if a download is active."""
         return model_name in self._active_downloads

backend/voicebox-server.spec

@@ -6,8 +6,13 @@ from PyInstaller.utils.hooks import copy_metadata
 datas = []
 hiddenimports = ['backend', 'backend.main', 'backend.config', 'backend.database', 'backend.models', 'backend.profiles', 'backend.history', 'backend.tts', 'backend.transcribe', 'backend.platform_detect', 'backend.backends', 'backend.backends.pytorch_backend', 'backend.utils.audio', 'backend.utils.cache', 'backend.utils.progress', 'backend.utils.hf_progress', 'backend.utils.validation', 'torch', 'transformers', 'fastapi', 'uvicorn', 'sqlalchemy', 'librosa', 'soundfile', 'qwen_tts', 'qwen_tts.inference', 'qwen_tts.inference.qwen3_tts_model', 'qwen_tts.inference.qwen3_tts_tokenizer', 'qwen_tts.core', 'qwen_tts.cli', 'pkg_resources.extern', 'backend.backends.mlx_backend', 'mlx', 'mlx.core', 'mlx.nn', 'mlx_audio', 'mlx_audio.tts', 'mlx_audio.stt']
 datas += collect_data_files('qwen_tts')
-datas += collect_data_files('mlx')
-datas += collect_data_files('mlx_audio')
+# Use collect_all (not collect_data_files) so native .dylib and .metallib
+# files are bundled as binaries, not data. Without this, MLX raises OSError
+# when loading Metal shaders inside the PyInstaller bundle.
+from PyInstaller.utils.hooks import collect_all as _collect_all
+_mlx_datas, _mlx_bins, _mlx_hidden = _collect_all('mlx')
+_mlxa_datas, _mlxa_bins, _mlxa_hidden = _collect_all('mlx_audio')
+datas += _mlx_datas + _mlxa_datas
 datas += copy_metadata('qwen-tts')
 hiddenimports += collect_submodules('qwen_tts')
 hiddenimports += collect_submodules('jaraco')
@@ -18,7 +23,7 @@ hiddenimports += collect_submodules('mlx_audio')
 a = Analysis(
     ['server.py'],
     pathex=[],
-    binaries=[],
+    binaries=_mlx_bins + _mlxa_bins,
     datas=datas,
     hiddenimports=hiddenimports,
     hookspath=[],

docs/plans/CUDA_BACKEND_SWAP.md

@@ -0,0 +1,581 @@
+# CUDA Backend Swap via Binary Replacement
+
+> Status: Plan | Target: v0.2.0 | Created: 2026-03-12
+
+## Problem
+
+The CUDA PyTorch backend binary is ~2.4 GB. GitHub Releases has a 2 GB asset limit. The current release ships CPU-only PyTorch on Windows and Intel Mac — NVIDIA GPU users get no acceleration from official releases. This is the #1 reported issue category (19 open issues).
+
+Users who want GPU today must clone the repo and run from source. That's not acceptable for a desktop app targeting non-technical users.
+
+## Solution
+
+Ship two backend binaries: a default CPU build (~150 MB) bundled with the app, and a downloadable CUDA build (~2.4 GB) hosted externally. When the user downloads the CUDA build, the app kills the current backend process, swaps in the CUDA binary, and relaunches — a backend-only restart. The frontend stays running, all UI state is preserved.
+
+No subprocesses. No HTTP protocol between processes. No port allocation. No provider manager. The backend is still one monolithic process — just a different binary.
+
+## Architecture
+
+### What Exists Today
+
+```
+Tauri App
+  ├── React Frontend (in-process webview)
+  └── voicebox-server (sidecar subprocess on :17493)
+        └── One PyInstaller binary: CPU PyTorch or MLX
+```
+
+**Sidecar lifecycle** (`tauri/src-tauri/src/main.rs`):
+- `start_server` command spawns `voicebox-server` sidecar (line 181)
+- Binary located at `tauri/src-tauri/binaries/voicebox-server-{platform-triple}`
+- Tauri resolves the sidecar name via `externalBin` in `tauri.conf.json` (line 16)
+- Waits up to 120s for "Uvicorn running" in stdout/stderr (line 286)
+- `stop_server` kills the process tree (line 466)
+
+**Frontend reconnection** (`app/src/lib/hooks/useServer.ts`):
+- Health check polls `GET /health` every 30 seconds
+- React Query cache retains data for 10 minutes after disconnect
+- All UI state (Zustand stores, form data, open tabs) survives disconnection
+- No active reconnect logic — just keeps polling until server responds
+
+This means a backend restart is mostly invisible to the frontend: it sees a few seconds of failed health checks, then the server comes back. The only risk is in-flight operations (generation, transcription) failing mid-request.
+
+### What Changes
+
+```
+Tauri App
+  ├── React Frontend (in-process webview)
+  └── voicebox-server (sidecar subprocess on :17493)
+        └── One of:
+            ├── voicebox-server-cpu     (bundled, ~150 MB)
+            └── voicebox-server-cuda    (downloaded, ~2.4 GB)
+```
+
+The CUDA binary is functionally identical to the CPU binary. Same FastAPI app, same endpoints, same code. The only difference is PyTorch is compiled with CUDA 12.1 support and the binary includes CUDA runtime libraries.
+
+The user downloads it once. On every subsequent app launch, Tauri checks which binary variant exists and spawns the appropriate one.
+
+## Implementation Plan
+
+### Phase 1: Build Infrastructure
+
+Build the CUDA binary in CI separately from the main release.
+
+#### 1a. CUDA PyInstaller Build
+
+Add a `build_binary_cuda.py` or parameterize the existing `build_binary.py`:
+
+```python
+# backend/build_binary.py — add flag
+def build_server(cuda=False):
+    args = [
+        'server.py',
+        '--onefile',
+        '--name', f'voicebox-server-{"cuda" if cuda else "cpu"}',
+    ]
+
+    if cuda:
+        args.extend([
+            '--hidden-import', 'torch.cuda',
+            '--hidden-import', 'torch.backends.cudnn',
+        ])
+    # ... rest of existing build
+```
+
+The `--onefile` flag is already used, which produces a single executable. This is important — `--onedir` would complicate the swap (replacing a directory vs a file).
+
+#### 1b. CI Workflow for CUDA Binary
+
+New workflow: `.github/workflows/build-cuda.yml`
+
+```yaml
+name: Build CUDA Provider
+on:
+  workflow_dispatch:
+  push:
+    tags: ["v*"]
+
+jobs:
+  build-cuda:
+    runs-on: windows-latest  # CUDA is Windows/Linux only
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with: { python-version: "3.12" }
+      - name: Install dependencies
+        run: |
+          pip install pyinstaller
+          pip install -r backend/requirements.txt
+          pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall
+      - name: Build CUDA binary
+        run: python backend/build_binary.py --cuda
+      - name: Split binary for GitHub Releases
+        run: |
+          python scripts/split_binary.py backend/dist/voicebox-server-cuda.exe \
+            --chunk-size 1900MB \
+            --output release-assets/
+      - name: Upload to R2
+        # Full binary to R2 (no size limit)
+        run: |
+          aws s3 cp backend/dist/voicebox-server-cuda.exe \
+            s3://voicebox-downloads/cuda/v${{ github.ref_name }}/voicebox-server-cuda.exe \
+            --endpoint-url ${{ secrets.R2_ENDPOINT }}
+      - name: Upload split parts to GitHub Release
+        # Split parts as GitHub Release assets (each <2 GB)
+        uses: softprops/action-gh-release@v1
+        with:
+          files: release-assets/*
+```
+
+Two distribution paths for redundancy:
+- **Cloudflare R2**: Full binary, direct download, no size limit.
+- **GitHub Releases**: Split into <2 GB chunks as fallback.
+
+#### 1c. Binary Splitting Script
+
+```python
+# scripts/split_binary.py
+"""Split a large binary into chunks for GitHub Releases."""
+import hashlib
+import argparse
+from pathlib import Path
+
+def split(input_path: Path, chunk_size: int, output_dir: Path):
+    output_dir.mkdir(parents=True, exist_ok=True)
+    data = input_path.read_bytes()
+
+    # Write SHA-256 of the complete file
+    sha256 = hashlib.sha256(data).hexdigest()
+    (output_dir / f"{input_path.stem}.sha256").write_text(
+        f"{sha256}  {input_path.name}\n"
+    )
+
+    # Split into chunks
+    parts = []
+    for i in range(0, len(data), chunk_size):
+        part_name = f"{input_path.stem}.part{len(parts):02d}{input_path.suffix}"
+        part_path = output_dir / part_name
+        part_path.write_bytes(data[i:i + chunk_size])
+        parts.append(part_name)
+
+    # Write manifest
+    (output_dir / f"{input_path.stem}.manifest").write_text(
+        "\n".join(parts) + "\n"
+    )
+
+    print(f"Split into {len(parts)} parts, SHA-256: {sha256}")
+```
+
+### Phase 2: Download & Assemble in App
+
+#### 2a. Backend Download Endpoint
+
+Add to `backend/main.py`:
+
+```python
+@app.post("/backend/download-cuda")
+async def download_cuda_backend():
+    """Download the CUDA backend binary."""
+    # Returns immediately, runs download in background
+    task = asyncio.create_task(_download_cuda_binary())
+    task.add_done_callback(lambda t: logger.error(f"CUDA download failed: {t.exception()}") if t.exception() else None)
+    return {"status": "downloading"}
+
+@app.get("/backend/cuda-status")
+async def cuda_status():
+    """Check if CUDA binary is available."""
+    cuda_path = _get_cuda_binary_path()
+    return {
+        "available": cuda_path is not None and cuda_path.exists(),
+        "active": _is_cuda_active(),
+        "download_progress": progress_manager.get_progress("cuda-backend"),
+    }
+```
+
+#### 2b. Download + Assemble + Verify Logic
+
+New file: `backend/cuda_download.py`
+
+Core logic:
+
+```python
+import hashlib
+from pathlib import Path
+from backend.config import get_data_dir
+from backend.utils.progress import get_progress_manager
+
+CUDA_DOWNLOAD_URL = "https://downloads.voicebox.sh/cuda/{version}/voicebox-server-cuda{ext}"
+CUDA_CHECKSUMS = {
+    # Populated per release
+    "0.2.0-windows": "sha256:abc123...",
+    "0.2.0-linux": "sha256:def456...",
+}
+
+def get_cuda_binary_dir() -> Path:
+    """Where CUDA binaries live. Inside the app's data directory."""
+    return get_data_dir() / "backends"
+
+def get_cuda_binary_path() -> Path | None:
+    """Return path to CUDA binary if it exists and is verified."""
+    d = get_cuda_binary_dir()
+    for name in ["voicebox-server-cuda.exe", "voicebox-server-cuda"]:
+        p = d / name
+        if p.exists():
+            return p
+    return None
+
+async def download_cuda_binary(version: str):
+    """Download, assemble (if split), and verify the CUDA binary."""
+    progress = get_progress_manager()
+    dest_dir = get_cuda_binary_dir()
+    dest_dir.mkdir(parents=True, exist_ok=True)
+
+    ext = ".exe" if sys.platform == "win32" else ""
+    url = CUDA_DOWNLOAD_URL.format(version=version, ext=ext)
+
+    # Download with progress tracking
+    temp_path = dest_dir / f"voicebox-server-cuda{ext}.download"
+    async with httpx.AsyncClient(follow_redirects=True) as client:
+        async with client.stream("GET", url) as response:
+            total = int(response.headers.get("content-length", 0))
+            downloaded = 0
+            with open(temp_path, "wb") as f:
+                async for chunk in response.aiter_bytes(chunk_size=1024 * 1024):
+                    f.write(chunk)
+                    downloaded += len(chunk)
+                    progress.update("cuda-backend", downloaded, total)
+
+    # Verify checksum
+    sha256 = hashlib.sha256(temp_path.read_bytes()).hexdigest()
+    expected = CUDA_CHECKSUMS.get(f"{version}-{sys.platform}")
+    if expected and not expected.endswith(sha256):
+        temp_path.unlink()
+        raise ValueError(f"Checksum mismatch: expected {expected}, got sha256:{sha256}")
+
+    # Atomic move into place
+    final_path = dest_dir / f"voicebox-server-cuda{ext}"
+    temp_path.rename(final_path)
+
+    # Make executable on Unix
+    if sys.platform != "win32":
+        final_path.chmod(0o755)
+
+    progress.complete("cuda-backend")
+```
+
+Key points:
+- Downloads to a `.download` temp file, verifies checksum, then atomically renames. No partial binaries left on crash.
+- Progress tracked via the existing `ProgressManager` so the frontend SSE system works unchanged.
+- CUDA binary lives in the **app data directory** (`data/backends/`), not alongside the app bundle. This avoids code-signing issues on macOS (though CUDA isn't relevant on macOS) and survives app updates.
+
+#### 2c. Reassembly from Split Parts (GitHub Releases Fallback)
+
+If the R2 download fails, fall back to downloading split parts from GitHub Releases:
+
+```python
+async def download_cuda_from_github(version: str):
+    """Fallback: download split parts from GitHub Releases, reassemble."""
+    base_url = f"https://github.com/jamiepine/voicebox/releases/download/v{version}"
+
+    # Get manifest
+    manifest_url = f"{base_url}/voicebox-server-cuda.manifest"
+    async with httpx.AsyncClient(follow_redirects=True) as client:
+        manifest = (await client.get(manifest_url)).text
+        parts = [p.strip() for p in manifest.strip().splitlines()]
+
+        # Download checksum
+        sha256_url = f"{base_url}/voicebox-server-cuda.sha256"
+        expected_sha = (await client.get(sha256_url)).text.split()[0]
+
+        # Download parts
+        dest_dir = get_cuda_binary_dir()
+        dest_dir.mkdir(parents=True, exist_ok=True)
+        temp_path = dest_dir / "voicebox-server-cuda.exe.download"
+
+        total_downloaded = 0
+        with open(temp_path, "wb") as f:
+            for i, part_name in enumerate(parts):
+                part_url = f"{base_url}/{part_name}"
+                async with client.stream("GET", part_url) as response:
+                    async for chunk in response.aiter_bytes(chunk_size=1024 * 1024):
+                        f.write(chunk)
+                        total_downloaded += len(chunk)
+                        get_progress_manager().update(
+                            "cuda-backend", total_downloaded, None,
+                            message=f"Downloading part {i+1}/{len(parts)}"
+                        )
+
+    # Verify reassembled file
+    sha256 = hashlib.sha256(temp_path.read_bytes()).hexdigest()
+    if sha256 != expected_sha:
+        temp_path.unlink()
+        raise ValueError(f"Checksum mismatch after reassembly")
+
+    final_path = dest_dir / "voicebox-server-cuda.exe"
+    temp_path.rename(final_path)
+    get_progress_manager().complete("cuda-backend")
+```
+
+### Phase 3: Backend Restart (The Swap)
+
+This is the core of the feature: kill the CPU backend, launch the CUDA backend, frontend reconnects automatically.
+
+#### 3a. New Tauri Command: `restart_server`
+
+Add to `tauri/src-tauri/src/main.rs`:
+
+```rust
+#[command]
+async fn restart_server(
+    app: tauri::AppHandle,
+    state: State<'_, ServerState>,
+    use_cuda: Option<bool>,
+) -> Result<String, String> {
+    println!("restart_server: use_cuda={:?}", use_cuda);
+
+    // 1. Stop the current server
+    stop_server(state.clone()).await?;
+
+    // 2. Brief wait for port release
+    tokio::time::sleep(tokio::time::Duration::from_millis(500)).await;
+
+    // 3. Start with the appropriate binary
+    // The start_server logic needs to check for CUDA binary
+    start_server(app, state, None).await
+}
+```
+
+#### 3b. Modify `start_server` to Prefer CUDA Binary
+
+The existing `start_server` uses `app.shell().sidecar("voicebox-server")` which resolves via Tauri's `externalBin` config. For the CUDA binary (which lives in the data directory, not the app bundle), we need an alternative launch path.
+
+Modify `start_server` in `main.rs`:
+
+```rust
+// After the existing sidecar logic, before spawning:
+
+// Check for CUDA binary in data directory
+let cuda_binary = data_dir.join("backends")
+    .join(if cfg!(windows) { "voicebox-server-cuda.exe" } else { "voicebox-server-cuda" });
+
+let (mut rx, child) = if cuda_binary.exists() {
+    println!("Found CUDA backend binary at {:?}", cuda_binary);
+
+    // Launch CUDA binary directly (not as Tauri sidecar)
+    let mut cmd = app.shell().command(cuda_binary.to_str().unwrap());
+    cmd = cmd.args([
+        "--data-dir",
+        data_dir.to_str().ok_or("Invalid data dir path")?,
+        "--port",
+        &SERVER_PORT.to_string(),
+    ]);
+    if remote.unwrap_or(false) {
+        cmd = cmd.args(["--host", "0.0.0.0"]);
+    }
+    cmd.spawn().map_err(|e| format!("Failed to spawn CUDA backend: {}", e))?
+} else {
+    // Existing sidecar launch (CPU binary bundled with app)
+    sidecar.spawn().map_err(|e| format!("Failed to spawn: {}", e))?
+};
+```
+
+Key decisions:
+- CUDA binary is launched via `app.shell().command()` (arbitrary path), not `app.shell().sidecar()` (bundled path). Tauri's sidecar system only resolves binaries within the app bundle.
+- The CUDA binary gets the same args (`--data-dir`, `--port`) as the CPU binary. It's the same `server.py` entry point.
+- Preference: if CUDA binary exists, use it. Otherwise fall back to bundled CPU. No user configuration needed.
+
+#### 3c. Frontend: Trigger Restart After Download
+
+Add to the platform lifecycle interface (`app/src/platform/types.ts`):
+
+```typescript
+interface PlatformLifecycle {
+    startServer(remote?: boolean): Promise<string>;
+    stopServer(): Promise<void>;
+    restartServer(useCuda?: boolean): Promise<string>;  // new
+    // ...
+}
+```
+
+Implement in `tauri/src/platform/lifecycle.ts`:
+
+```typescript
+async restartServer(useCuda?: boolean): Promise<string> {
+    const result = await invoke<string>('restart_server', { useCuda });
+    this.onServerReady?.();
+    return result;
+}
+```
+
+#### 3d. Frontend: GPU Settings UI
+
+Add a section to the Server Settings page (or Model Management). Minimal UI:
+
+```
+┌─────────────────────────────────────────────┐
+│  GPU Acceleration                           │
+│                                             │
+│  Status: CPU only (no CUDA backend)         │
+│                                             │
+│  [Download CUDA Backend (2.4 GB)]           │
+│                                             │
+│  Requires an NVIDIA GPU with 4+ GB VRAM.    │
+│  The app will restart its backend process   │
+│  after download. Your work is preserved.    │
+└─────────────────────────────────────────────┘
+```
+
+After download:
+
+```
+┌─────────────────────────────────────────────┐
+│  GPU Acceleration                           │
+│                                             │
+│  Status: ✓ CUDA backend active (RTX 4090)   │
+│                                             │
+│  [Switch to CPU]     [Delete CUDA Backend]  │
+└─────────────────────────────────────────────┘
+```
+
+#### 3e. Frontend: Reconnection During Restart
+
+The current health poll interval is 30 seconds — too slow for a restart UX. During a restart, temporarily increase polling:
+
+```typescript
+// In the component that triggers restart:
+const restart = async () => {
+    setRestarting(true);
+    try {
+        await platform.lifecycle.restartServer(true);
+    } catch (e) {
+        // Frontend will show "reconnecting" state
+    }
+    // Aggressively poll until health check succeeds
+    const interval = setInterval(async () => {
+        try {
+            await apiClient.getHealth();
+            clearInterval(interval);
+            setRestarting(false);
+            queryClient.invalidateQueries(); // Refresh all data
+        } catch {}
+    }, 1000); // Poll every 1s during restart
+    // Safety timeout
+    setTimeout(() => clearInterval(interval), 30000);
+};
+```
+
+### Phase 4: Auto-Detection on Startup
+
+No user action needed on subsequent launches. The preference logic in `start_server` (Phase 3b) handles this:
+
+1. App launches → `start_server` called
+2. Check `data/backends/voicebox-server-cuda{.exe}`
+3. If exists → launch CUDA binary
+4. If not → launch bundled CPU binary
+
+The user downloads CUDA once, and every future app launch (including after updates) uses it automatically. The CUDA binary lives in the data directory, not the app bundle, so app updates don't overwrite it.
+
+### Phase 5: Handling Version Mismatches
+
+When the app updates but the CUDA binary is from an older version, the API might be incompatible. Handle this by:
+
+1. Add `--version` flag to `server.py`:
+
+```python
+parser.add_argument("--version", action="store_true")
+# If invoked with --version, print version and exit
+if args.version:
+    from backend import __version__
+    print(f"voicebox-server {__version__}")
+    sys.exit(0)
+```
+
+2. In `start_server` (Rust), before launching the CUDA binary:
+
+```rust
+// Quick version check
+let version_output = std::process::Command::new(cuda_binary.to_str().unwrap())
+    .arg("--version")
+    .output();
+
+match version_output {
+    Ok(output) => {
+        let version = String::from_utf8_lossy(&output.stdout);
+        let app_version = env!("CARGO_PKG_VERSION");
+        if !version.contains(app_version) {
+            println!("CUDA binary version mismatch (app: {}, cuda: {}), falling back to CPU",
+                app_version, version.trim());
+            // Fall through to CPU sidecar launch
+        }
+    }
+    Err(_) => {
+        println!("Failed to check CUDA binary version, falling back to CPU");
+    }
+}
+```
+
+3. Frontend shows a notification: "Your GPU backend needs an update. [Download latest] or [Use CPU for now]"
+
+## Files Changed
+
+### New Files
+
+| File | Purpose |
+|------|---------|
+| `backend/cuda_download.py` | Download, reassemble, verify CUDA binary |
+| `scripts/split_binary.py` | Split binary into <2 GB chunks for GitHub Releases |
+| `.github/workflows/build-cuda.yml` | CI: build + upload CUDA binary |
+
+### Modified Files
+
+| File | Change |
+|------|--------|
+| `tauri/src-tauri/src/main.rs` | Add `restart_server` command, modify `start_server` to check for CUDA binary in data dir |
+| `backend/server.py` | Add `--version` flag |
+| `backend/main.py` | Add `/backend/download-cuda`, `/backend/cuda-status`, `/backend/progress/cuda-backend` endpoints |
+| `backend/build_binary.py` | Accept `--cuda` flag to build CUDA variant |
+| `app/src/platform/types.ts` | Add `restartServer` to lifecycle interface |
+| `tauri/src/platform/lifecycle.ts` | Implement `restartServer` |
+| `app/src/components/ServerSettings/` | New GPU acceleration section |
+| `.github/workflows/release.yml` | Trigger CUDA build workflow on tag |
+
+### NOT Changed
+
+| File | Why |
+|------|-----|
+| `backend/backends/__init__.py` | No changes to the TTSBackend singleton or factory. CUDA binary runs the same code. |
+| `backend/backends/pytorch_backend.py` | Already detects CUDA at runtime (line 28-49). No changes needed. |
+| `app/src/lib/api/client.ts` | API is identical between CPU and CUDA backends. |
+| `app/src/lib/hooks/useGenerationForm.ts` | Generation flow is unchanged. |
+
+## What This Doesn't Solve
+
+- **Multi-model support** — This is purely about GPU acceleration. LuxTTS, Chatterbox, etc. need the in-process model registry, which is an independent workstream.
+- **AMD GPU support** — DirectML/ROCm needs a different PyTorch build. Same pattern applies (another binary variant) but deferred.
+- **Linux CUDA** — Same approach works, just another CI matrix entry. Can be added in the same release or shortly after.
+- **Remote server mode** — Users who want to run TTS on a different machine still need the external provider architecture. Separate concern.
+
+## What This DOES Solve
+
+- **19 "GPU not detected" issues** — Users download the CUDA backend, restart, GPU works.
+- **2 GB GitHub Release limit** — Binary splitting + R2 hosting.
+- **Update burden** — App updates don't re-download the 2.4 GB CUDA binary. It persists in the data directory.
+- **First-run experience** — App works immediately on CPU. GPU is an optional enhancement, not a setup blocker.
+
+## Rollout Plan
+
+1. Build and test CUDA binary locally on Windows with an NVIDIA GPU.
+2. Set up R2 bucket at `downloads.voicebox.sh/cuda/`.
+3. Ship the backend restart + download UI in v0.2.0.
+4. Announce: "GPU acceleration is here — one click in Settings."
+
+## Risks
+
+| Risk | Mitigation |
+|------|-----------|
+| CUDA binary doesn't work on some GPU/driver combos | `/health` endpoint reports GPU info. Fallback to CPU if CUDA init fails. Clear error message. |
+| Antivirus flags downloaded binary (Windows) | Code-sign the CUDA binary in CI. Document AV exceptions. |
+| Data dir CUDA binary survives app uninstall | Document in uninstall notes. Not a real problem — it's just a file. |
+| Version mismatch after app update | Version check on startup (Phase 5). Auto-fallback to CPU. Prompt to re-download. |
+| R2 downtime | GitHub Releases split-binary fallback. |
+| Download interrupted | Temp file with `.download` extension. Atomic rename on completion. Resume not implemented in v1 — restart download from scratch. |

docs/plans/CUDA_BACKEND_SWAP_FINAL.md

@@ -0,0 +1,133 @@
+# CUDA Backend Swap — Implementation Summary
+
+> Status: **Complete** | Branch: `feat/cuda-backend-swap` | Created: 2026-03-12
+
+## What This Is
+
+A standalone feature that lets users download a CUDA-enabled backend binary (~2.4 GB) and swap it in via a backend-only restart. The frontend stays running, all UI state is preserved. This solves the #1 user pain point: 19 open issues about "GPU not detected" caused by GitHub's 2 GB release asset limit preventing CUDA binaries from shipping in official releases.
+
+## How It Works
+
+```
+User clicks "Download CUDA Backend" in Settings
+  → Backend fetches manifest from GitHub Releases
+  → Downloads split parts (<2 GB each), concatenates them
+  → SHA-256 integrity check on reassembled binary
+  → Binary placed in {app_data_dir}/backends/voicebox-server-cuda
+  → User clicks "Switch to CUDA Backend"
+  → Tauri kills CPU process, launches CUDA binary, frontend reconnects
+  → On all future app launches, CUDA binary is auto-detected and used
+```
+
+The CUDA binary is functionally identical to the CPU binary — same FastAPI app, same endpoints, same code. The only difference is PyTorch compiled with CUDA 12.1 and bundled CUDA runtime libraries.
+
+## Architecture Decisions
+
+**Backend-only restart, not full app restart.** The Tauri shell kills the current `voicebox-server` process, waits 1 second for port release, and spawns the new binary. The React frontend stays running. Health polling detects the new backend within seconds.
+
+**No provider/subprocess architecture.** This is explicitly not the PR #33 approach (10K+ lines, 136 files, 22 bugs). One process at a time. The CUDA binary replaces the CPU binary — it doesn't run alongside it.
+
+**Data directory, not app bundle.** The CUDA binary lives in `{app_data_dir}/backends/`, which persists across app updates and avoids code-signing issues. The bundled CPU binary in the app bundle is untouched.
+
+**Version mismatch protection.** On startup, Rust runs `voicebox-server-cuda --version` and compares to the app version from `tauri.conf.json`. If they don't match (e.g., after an app update), it falls back to the bundled CPU binary silently.
+
+**GitHub Releases distribution.** The CUDA binary is split into <2 GB chunks (GitHub's asset limit) via `scripts/split_binary.py`. The app downloads a manifest, fetches each part, concatenates them, and runs a SHA-256 integrity check to verify reassembly. No external hosting needed.
+
+## Files Changed
+
+### New Files
+
+| File | Lines | Purpose |
+|------|-------|---------|
+| `backend/cuda_download.py` | ~190 | Download split parts from GitHub Releases, reassemble, verify integrity |
+| `scripts/split_binary.py` | ~80 | Split large binary into <2 GB chunks with SHA-256 manifest |
+| `.github/workflows/build-cuda.yml` | ~70 | CI workflow: build CUDA binary, split, upload to GitHub Releases |
+| `app/src/components/ServerSettings/GpuAcceleration.tsx` | 371 | GPU Acceleration UI card (status, download, restart, delete) |
+| `docs/plans/CUDA_BACKEND_SWAP.md` | 581 | Original implementation plan (5 phases with code sketches) |
+| `docs/plans/CUDA_BACKEND_SWAP_FINAL.md` | this file | Final implementation summary |
+| `docs/plans/PROJECT_STATUS.md` | 462 | Full project triage (all PRs, issues, architecture) |
+| `docs/plans/PR33_CUDA_PROVIDER_REVIEW.md` | ~350 | Detailed code review of PR #33 (22 bugs documented) |
+
+### Modified Files
+
+| File | What Changed |
+|------|-------------|
+| `backend/build_binary.py` | Added `--cuda` flag, parameterized output binary name |
+| `backend/server.py` | Added `--version` flag, auto-detect backend variant from binary name (`VOICEBOX_BACKEND_VARIANT` env var) |
+| `backend/main.py` | 4 new endpoints (`/backend/cuda-status`, `/backend/download-cuda`, `/backend/cuda`, `/backend/cuda-progress`), health endpoint returns `backend_variant` |
+| `backend/models.py` | `HealthResponse` model: added `backend_variant` field |
+| `backend/requirements.txt` | Added `httpx>=0.27.0` for async HTTP downloads |
+| `tauri/src-tauri/src/main.rs` | `restart_server` command (stop → wait → start), `start_server` checks for CUDA binary in data dir and launches via `shell().command()`, version mismatch check |
+| `app/src/platform/types.ts` | `PlatformLifecycle.restartServer()` added |
+| `tauri/src/platform/lifecycle.ts` | `restartServer()` implementation via `invoke('restart_server')` |
+| `web/src/platform/lifecycle.ts` | `restartServer()` noop for web platform |
+| `app/src/lib/api/types.ts` | `CudaStatus`, `CudaDownloadProgress` interfaces; `HealthResponse` updated with `gpu_type`, `backend_type`, `backend_variant` |
+| `app/src/lib/api/client.ts` | `getCudaStatus()`, `downloadCudaBackend()`, `deleteCudaBackend()` methods |
+| `app/src/components/ServerTab/ServerTab.tsx` | Wired in `<GpuAcceleration />` component (Tauri-only) |
+
+## Backend API Endpoints
+
+| Method | Path | Purpose |
+|--------|------|---------|
+| `GET` | `/backend/cuda-status` | Returns `{ available, active, binary_path, downloading, download_progress }` |
+| `POST` | `/backend/download-cuda` | Starts background download; returns immediately. Track via SSE. |
+| `DELETE` | `/backend/cuda` | Deletes CUDA binary (blocked if CUDA is currently active) |
+| `GET` | `/backend/cuda-progress` | SSE stream of download progress (reuses existing `ProgressManager`) |
+
+The existing `GET /health` endpoint now returns two new fields:
+- `backend_type`: `"pytorch"` or `"mlx"` (existing detection)
+- `backend_variant`: `"cpu"` or `"cuda"` (set from `VOICEBOX_BACKEND_VARIANT` env var)
+
+## Frontend UI States
+
+The `GpuAcceleration` card in Server Settings handles these states:
+
+1. **Native GPU detected** (MPS, MLX, XPU, DirectML) — Shows info message, no download needed
+2. **No CUDA binary** — Download button with size estimate, description of requirements
+3. **Downloading** — SSE-driven progress bar with bytes/total and percentage
+4. **Downloaded, not active** — "Switch to CUDA Backend" button + "Remove" option
+5. **CUDA active** — Shows CUDA badge, "Switch to CPU Backend" button
+6. **Restarting** — Spinner with phase text, 1s health polling as safety net
+7. **Error** — Red error message with details
+
+### Key UX detail: switching to CPU
+
+Since `start_server` always prefers the CUDA binary if it exists on disk, "Switch to CPU" must delete the CUDA binary first, then restart. The user can re-download later. This avoids a persistent configuration mechanism (no new state to manage, no new config file, no DB column).
+
+## Rust: Server Lifecycle
+
+```
+start_server
+  ├── Check for CUDA binary at {data_dir}/backends/voicebox-server-cuda
+  ├── If found: run --version, compare to app version
+  │   ├── Match: launch via shell().command() with --data-dir, --port
+  │   └── Mismatch: log warning, fall through to CPU
+  └── Else: launch bundled sidecar via shell().sidecar()
+
+restart_server
+  ├── stop_server (kill process tree)
+  ├── wait 1 second for port release
+  └── start_server (auto-detects CUDA)
+```
+
+## What This Doesn't Cover
+
+- **AMD GPU / ROCm / DirectML binary** — Same pattern, different PyTorch build. Future PR.
+- **Linux CUDA** — Same approach, just another CI matrix entry. Can ship same release.
+- **Multi-model support** — LuxTTS, Chatterbox, etc. are a separate architectural concern (in-process model registry). Independent of binary variant.
+- **Download resume** — If download is interrupted, it restarts from scratch. Acceptable for v1.
+- **Remote server CUDA** — Users running voicebox-server on a remote machine manage their own binaries. This feature is for the desktop app.
+
+## Testing Checklist
+
+- [ ] Build CUDA binary locally with `python backend/build_binary.py --cuda`
+- [ ] `voicebox-server-cuda --version` prints correct version
+- [ ] Place CUDA binary in `{data_dir}/backends/`, launch app → auto-detects and uses it
+- [ ] Version mismatch: rename binary to have wrong version → falls back to CPU
+- [ ] Frontend: GpuAcceleration card shows correct state for CPU, CUDA available, CUDA active
+- [ ] Download flow: POST triggers download, SSE progress works, completion updates status
+- [ ] Switch to CUDA: restart works, health endpoint shows `backend_variant: "cuda"`
+- [ ] Switch to CPU: deletes binary, restarts, health shows `backend_variant: "cpu"`
+- [ ] Delete CUDA while active: returns 409 error
+- [ ] Split binary script: `python scripts/split_binary.py` creates manifest + parts + sha256
+- [ ] Native GPU (macOS MPS): shows info message, no download section

docs/plans/PR33_CUDA_PROVIDER_REVIEW.md

@@ -0,0 +1,500 @@
+# PR #33 — CUDA Provider System Review
+
+> Branch: `external-provider-binaries` | Created: 2026-02-01 | 34 commits, 136 files, +10,266 lines
+> Reviewed: 2026-03-12
+
+---
+
+## The Problem
+
+The CUDA PyTorch binary is ~2.4 GB. GitHub Releases has a 2 GB artifact limit. This means:
+
+- Windows/Linux users with NVIDIA GPUs cannot get GPU acceleration from official releases
+- 19 open issues about "GPU not detected" — the single most reported problem category
+- Users who want GPU must clone the repo and run from source
+- Every app update forces re-download of the entire binary
+
+This is the #1 user pain point by volume.
+
+---
+
+## What PR #33 Does
+
+Splits the monolithic Voicebox binary into two layers:
+
+```
+┌──────────────────────────────────────┐
+│  Main App (~150MB Win/Lin, ~300 Mac) │
+│  Tauri + React + FastAPI + Whisper   │
+│  No PyTorch. MLX bundled on macOS.   │
+├──────────────────────────────────────┤
+│           HTTP (localhost)           │
+├──────────────────────────────────────┤
+│  Provider Binary (downloaded later)  │
+│  PyTorch CPU (~300MB)                │
+│  PyTorch CUDA (~2.4GB)              │
+│  Hosted on Cloudflare R2             │
+└──────────────────────────────────────┘
+```
+
+### New Backend Code
+
+| File | Purpose |
+|------|---------|
+| `backend/providers/__init__.py` (327 lines) | `ProviderManager` — lifecycle management, subprocess spawning, port allocation |
+| `backend/providers/base.py` (97 lines) | `TTSProvider` Protocol definition |
+| `backend/providers/bundled.py` (144 lines) | `BundledProvider` — wraps existing MLX/PyTorch backends for the new interface |
+| `backend/providers/local.py` (191 lines) | `LocalProvider` — HTTP client that talks to external provider processes |
+| `backend/providers/installer.py` (262 lines) | Download, extract, delete provider binaries |
+| `backend/providers/types.py` (34 lines) | `ProviderType` enum, `ProviderInfo` dataclass |
+| `backend/providers/checksums.py` (11 lines) | Checksum dict (currently empty) |
+
+### Provider Servers (Standalone Executables)
+
+| File | Purpose |
+|------|---------|
+| `providers/pytorch-cpu/main.py` (238 lines) | FastAPI server wrapping PyTorch CPU inference |
+| `providers/pytorch-cuda/main.py` (238 lines) | FastAPI server wrapping PyTorch CUDA inference |
+| `providers/pytorch-*/build.py` | PyInstaller build scripts |
+| `providers/pytorch-*/requirements.txt` | Isolated dependencies |
+
+### Frontend
+
+| File | Purpose |
+|------|---------|
+| `app/src/components/ServerSettings/ProviderSettings.tsx` (400 lines) | Provider download/start/stop/delete UI |
+
+### Also Included (Scope Creep)
+
+The PR bundles several unrelated changes that inflate the diff:
+
+- `docs2/` — Entire documentation site rewrite (Fumadocs migration, ~3000 lines)
+- `Dockerfile`, `Dockerfile.cuda`, `docker-compose.yml` — Docker support
+- `landing/` — Banner removal
+- UI refactors in Stories, History, Voice Profiles, Audio tab
+- Linux audio capture module
+- Various dependency bumps
+
+---
+
+## Bug Report
+
+### Critical — Will Crash at Runtime
+
+#### C1. Provider `generate` endpoint can't parse requests
+
+**`providers/pytorch-cpu/main.py:91-97`** (same in pytorch-cuda)
+
+```python
+@app.post("/tts/generate")
+async def generate(
+    text: str,
+    voice_prompt: dict,
+    language: str = "auto",
+    seed: int = None,
+    model_size: str = "1.7B"
+):
+```
+
+Parameters declared as function arguments. FastAPI interprets these as **query parameters**, not JSON body. But `LocalProvider.generate()` sends a JSON body via `httpx`:
+
+```python
+# backend/providers/local.py:33-40
+response = await self.client.post("/tts/generate", json={
+    "text": text,
+    "voice_prompt": voice_prompt,
+    ...
+})
+```
+
+**Result:** Every generation call to an external provider returns HTTP 422 (Validation Error). The generation path is completely broken for external providers.
+
+**Fix:** Use a Pydantic request body model:
+```python
+class GenerateRequest(BaseModel):
+    text: str
+    voice_prompt: dict
+    language: str = "auto"
+    seed: Optional[int] = None
+    model_size: str = "1.7B"
+
+@app.post("/tts/generate")
+async def generate(data: GenerateRequest):
+```
+
+#### C2. Timeout error handler references undefined variables
+
+**`backend/providers/__init__.py:82-90`**
+
+```python
+stdout_content = ""
+stderr_content = ""
+# ... threads write to stdout_queue / stderr_queue ...
+except TimeoutError:
+    while not stdout_queue.empty():
+        stdout_lines.append(stdout_queue.get_nowait())  # NameError
+    while not stderr_queue.empty():
+        stderr_lines.append(stderr_queue.get_nowait())  # NameError
+```
+
+`stdout_lines` and `stderr_lines` are never defined. Every provider startup timeout will throw `NameError`, masking the real failure cause. Then `stdout_content` and `stderr_content` are logged but they're still empty strings — the queue data is never assigned back.
+
+#### C3. Sync `get_tts_model()` ignores external provider in async context
+
+**`backend/tts.py:15-29`**
+
+```python
+def get_tts_model():
+    manager = get_provider_manager()
+    loop = asyncio.get_event_loop()
+    if loop.is_running():
+        # We're in an async context, but can't await here
+        return manager._get_default_provider()
+```
+
+FastAPI routes are async. This function is called from several code paths during generation. In async context it **always returns the bundled provider**, ignoring whatever external provider the user selected. The user downloads and starts a CUDA provider, but generation still runs on CPU.
+
+### Critical — Security
+
+#### C4. Path traversal via `tarfile.extractall()` (CVE-2007-4559)
+
+**`backend/providers/installer.py:115-118`**
+
+```python
+with tarfile.open(archive_path, 'r:gz') as tar_ref:
+    tar_ref.extractall(providers_dir)
+```
+
+No member path filtering. A crafted `.tar.gz` from a compromised CDN can write files anywhere on disk via `../` entries. Python 3.12+ emits a deprecation warning for exactly this pattern.
+
+**Fix:**
+```python
+tar_ref.extractall(providers_dir, filter='data')  # Python 3.12+
+```
+
+Or manually validate each member:
+```python
+for member in tar_ref.getmembers():
+    member_path = os.path.join(providers_dir, member.name)
+    if not os.path.commonpath([providers_dir, member_path]).startswith(str(providers_dir)):
+        raise ValueError(f"Path traversal attempt: {member.name}")
+tar_ref.extractall(providers_dir)
+```
+
+#### C5. No checksum verification on downloaded binaries
+
+**`backend/providers/checksums.py`**
+
+```python
+PROVIDER_CHECKSUMS = {}
+```
+
+Empty dict. `download_provider()` in `installer.py` never calls any verification function. Downloaded binaries are `chmod 0o755`'d and executed without integrity checks. A MitM or CDN compromise delivers arbitrary code.
+
+**Fix:** Populate checksums per release. Verify SHA-256 after download before extraction:
+```python
+import hashlib
+sha256 = hashlib.sha256(archive_path.read_bytes()).hexdigest()
+if sha256 != expected:
+    archive_path.unlink()
+    raise ValueError(f"Checksum mismatch for {provider_type}")
+```
+
+#### C6. Provider servers have no authentication
+
+**`providers/pytorch-cpu/main.py:18-23`**
+
+```python
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    ...
+)
+```
+
+Zero auth. Any local process — including browser JavaScript via localhost — can send requests to the provider on its ephemeral port. Port is discoverable by scanning.
+
+**Fix:** Generate a random token in the parent process, pass via environment variable to the child, validate in middleware:
+```python
+# Parent (ProviderManager)
+token = secrets.token_urlsafe(32)
+env = {**os.environ, "VOICEBOX_PROVIDER_TOKEN": token}
+process = subprocess.Popen([...], env=env, ...)
+
+# Child (provider server)
+EXPECTED_TOKEN = os.environ.get("VOICEBOX_PROVIDER_TOKEN")
+
+@app.middleware("http")
+async def verify_token(request, call_next):
+    if request.headers.get("X-Provider-Token") != EXPECTED_TOKEN:
+        return JSONResponse(status_code=403, content={"error": "unauthorized"})
+    return await call_next(request)
+```
+
+### Major — Will Cause Problems in Production
+
+#### M1. Leaked file handles on subprocess stdout/stderr
+
+**`backend/providers/__init__.py:68-73`**
+
+```python
+process = subprocess.Popen(
+    [...],
+    stdout=open(stdout_log, 'w'),   # leaked handle
+    stderr=open(stderr_log, 'w'),   # leaked handle
+)
+```
+
+File handles passed directly from `open()` without storing references. They close on GC, not deterministically. On Windows the log files stay locked and unreadable until the process exits.
+
+**Fix:**
+```python
+stdout_fh = open(stdout_log, 'w')
+stderr_fh = open(stderr_log, 'w')
+try:
+    process = subprocess.Popen([...], stdout=stdout_fh, stderr=stderr_fh)
+finally:
+    stdout_fh.close()
+    stderr_fh.close()
+```
+
+#### M2. No subprocess crash detection or recovery
+
+**`backend/providers/__init__.py:56-110`**
+
+Once `start_provider()` succeeds, the `Popen` object is stored but never polled. If the provider process crashes mid-session:
+- `LocalProvider` HTTP calls fail with `httpx.ConnectError`
+- No auto-restart
+- No health-check loop
+- User sees cryptic "connection refused" errors
+- Must manually restart provider from UI
+
+**Fix:** Background asyncio task that polls `process.poll()` every few seconds. On crash, update provider status and optionally auto-restart:
+```python
+async def _watch_provider_process(self):
+    while self._provider_process and self._provider_process.poll() is None:
+        await asyncio.sleep(5)
+    if self._provider_process and self._provider_process.returncode != 0:
+        logger.error(f"Provider crashed with code {self._provider_process.returncode}")
+        self.active_provider = self._default_provider
+        # Notify frontend via next health check
+```
+
+#### M3. Port allocation race condition (TOCTOU)
+
+**`backend/providers/__init__.py:145-149`**
+
+```python
+def _get_free_port(self) -> int:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(('', 0))
+        return s.getsockname()[1]
+    # Socket closed here — port is free but unprotected
+```
+
+Between this function returning and the provider process binding, another process can claim the port. On busy systems this causes "address already in use" failures.
+
+**Fix options:**
+- Pass the socket fd to the child process (complex, platform-specific)
+- Retry with a new port on bind failure (simplest)
+- Use a fixed port range and try sequentially
+
+#### M4. `delete_provider()` leaves hundreds of MB behind
+
+**`backend/providers/installer.py:155-168`**
+
+```python
+provider_path.unlink()  # Deletes just the executable
+```
+
+PyInstaller `--onedir` produces a directory with the executable plus all shared libraries. `unlink()` only removes the binary file, leaving behind hundreds of MB of `.so`/`.dll`/`.dylib` files.
+
+**Fix:**
+```python
+provider_dir = provider_path.parent
+shutil.rmtree(provider_dir)
+```
+
+#### M5. `LocalProvider.combine_voice_prompts()` bypasses the provider
+
+**`backend/providers/local.py:68-88`**
+
+This method imports from `..utils.audio` and processes locally instead of sending to the provider server. If the user chose an external provider because they lack local dependencies (e.g., no PyTorch on the machine), this will crash with `ImportError`.
+
+#### M6. Download errors silently swallowed
+
+**`backend/main.py:1640`**
+
+```python
+asyncio.create_task(download_provider(provider_type))
+```
+
+Fire-and-forget. If the download fails, the exception is logged as "Task exception was never retrieved." The frontend SSE progress stream may hang forever showing "downloading" without the error.
+
+**Fix:** Store the task, add an error callback:
+```python
+task = asyncio.create_task(download_provider(provider_type))
+task.add_done_callback(lambda t: t.exception() if not t.cancelled() else None)
+```
+And propagate errors through the progress manager so the SSE stream surfaces them.
+
+#### M7. `LocalProvider.is_loaded()` always returns `True`
+
+**`backend/providers/local.py:105-108`**
+
+```python
+def is_loaded(self) -> bool:
+    return True  # Return True optimistically
+```
+
+Health/status checks always report the model as loaded for external providers, even when the provider hasn't loaded anything yet. This breaks the "download model if not cached" logic in the generation flow.
+
+#### M8. `instruct` parameter silently dropped
+
+**`backend/providers/local.py:33-40`**
+
+The `generate()` method accepts `instruct` but never includes it in the JSON payload. The provider server also hardcodes `instruct=None`. Delivery instructions silently do nothing for external providers.
+
+### Minor
+
+| # | Issue | Location |
+|---|-------|----------|
+| m1 | `pytorch-cpu/main.py` and `pytorch-cuda/main.py` are 95% identical | Both files | 
+| m2 | `build.py` scripts also nearly identical | Both build files |
+| m3 | `navigator.platform` is deprecated | `ProviderSettings.tsx:20-23` |
+| m4 | `console.log('currentProvider', ...)` left in | `ProviderSettings.tsx:151` |
+| m5 | `ProviderType` enum defined but never used for validation | `types.py:10-15` |
+| m6 | `list_installed()` reimplements platform detection | `__init__.py:129-143` |
+| m7 | New `httpx.AsyncClient` created per health poll iteration | `__init__.py:151-165` |
+| m8 | `load_model_async()` only stores size, doesn't actually preload | `local.py:95-99` |
+
+---
+
+## Scope Creep
+
+The PR should be split. These are independent changes bundled in:
+
+| Change | Lines | Should Be Separate PR |
+|--------|-------|-----------------------|
+| `docs2/` site rewrite | ~3000 | Yes |
+| Docker support (Dockerfile, compose, docs) | ~600 | Yes — overlaps with PR #161 |
+| Landing page banner removal | ~30 | Yes |
+| UI refactors (Stories, History, Voices, Audio) | ~400 | Yes |
+| Linux audio capture module | ~10 | Yes |
+| Dependency bumps | ~100 | Yes |
+
+**Core provider system** (the actual feature) is ~2500 lines across backend + frontend + provider servers. That's the reviewable scope.
+
+---
+
+## What's Well-Designed
+
+These parts should survive any rewrite:
+
+1. **`TTSProvider` Protocol** (`base.py`) — Structural typing via `@runtime_checkable Protocol`. Right pattern. Comprehensive interface.
+
+2. **`BundledProvider` / `LocalProvider` split** — Clean separation between in-process and HTTP-based inference. The wrapper pattern in `BundledProvider` correctly delegates to existing `TTSBackend`.
+
+3. **R2 distribution strategy** — Provider binaries on Cloudflare R2, main app on GitHub Releases. Correct solution to the 2 GB limit.
+
+4. **Progress tracking** — SSE-based download progress integrated with the existing `ProgressManager`. Good UX.
+
+5. **Subprocess log files** — Writing provider stdout/stderr to log files in the data directory is pragmatic and debuggable.
+
+6. **Frontend `ProviderSettings.tsx`** — Clean component structure. Proper loading/disabled states, confirmation dialogs, platform-aware visibility.
+
+7. **CI split** — Separate `build-providers` and `release` jobs. Providers built and uploaded to R2 independently.
+
+---
+
+## Options for Moving Forward
+
+### Option A — Fix and Slim PR #33
+
+Strip the PR down to just the provider system (~2500 lines). Fix the 5 critical and 8 major bugs. Rebase onto current `main`.
+
+**Effort:** ~2-3 days focused work
+**Pros:** Full auto-managed provider lifecycle. Foundation for multi-model.
+**Cons:** Still complex. Process management is inherently fragile cross-platform.
+
+### Option B — Manual External Server Mode
+
+Skip subprocess management entirely. Ship a "Connect to External Server" feature:
+
+1. User downloads CUDA provider zip from `downloads.voicebox.sh`
+2. User runs it manually (`./tts-provider-pytorch-cuda --port 8100`)
+3. In Voicebox UI: paste `http://localhost:8100` as the TTS server URL
+4. Voicebox routes generation to that URL via `LocalProvider`
+
+This reuses `LocalProvider` from PR #33 but removes:
+- `ProviderManager` subprocess spawning (the buggiest part)
+- `installer.py` download/extract logic (the security risks)
+- Port allocation (user picks the port)
+- Process lifecycle management (user's responsibility)
+
+**Effort:** ~1 day. `LocalProvider` + a URL input field + health check.
+**Pros:** Simple, reliable, no process management bugs, no security surface.
+**Cons:** Manual setup. Not seamless. But CUDA users are already technical (they run from source today).
+
+### Option C — Hybrid (Recommended)
+
+Ship Option B first as v0.2.0. Then iterate toward auto-management:
+
+**Phase 1 (v0.2.0):** Manual external server mode
+- `LocalProvider` HTTP client (from PR #33, with the 422 bug fixed)
+- Server URL input in Settings
+- Health indicator
+- CUDA provider published as standalone zip on R2
+- One page of docs: "download, unzip, run, paste URL"
+
+**Phase 2 (v0.2.x):** Auto-download + auto-start
+- `installer.py` with checksum verification and safe extraction
+- `ProviderManager` subprocess spawning with crash detection
+- Provider settings UI with download/start/stop buttons
+
+**Phase 3 (v0.3.0):** Multi-model providers
+- Provider per model family (not just per hardware)
+- LuxTTS provider, Chatterbox provider, etc.
+- Provider marketplace / registry
+
+This gets CUDA into users' hands immediately (Phase 1 is ~1 day) while building toward the full vision incrementally. Each phase is independently shippable and testable.
+
+### Option D — GitHub Workaround
+
+Avoid the provider architecture entirely. Host CUDA binaries on R2 and add a download link in the app that opens the user's browser. User downloads the full monolithic CUDA build, replaces their existing install.
+
+**Effort:** Minimal — just hosting + a link.
+**Pros:** Zero architecture changes.
+**Cons:** Doesn't solve: multi-model, independent app updates, or the re-download-everything-on-update problem. Kicks the can.
+
+---
+
+## Recommendation
+
+**Option C (Hybrid)** is the strongest path. Specifically:
+
+1. **Now:** Close PR #33 as-is. It's too large, too buggy, and too stale to salvage as a single merge.
+
+2. **Extract:** Cherry-pick the good parts into small focused PRs:
+   - PR: `TTSProvider` Protocol + `BundledProvider` + `LocalProvider` (the abstractions)
+   - PR: Provider settings UI (the frontend)
+   - PR: `installer.py` + checksums (the download system)
+   - PR: CI changes for R2 upload (the distribution)
+
+3. **Ship Phase 1:** Manual external server mode. One small PR. Unblocks every CUDA user immediately.
+
+4. **Iterate:** Layer in auto-management once the manual mode is proven stable.
+
+The critical bugs in PR #33 (C1-C6) are all fixable, but the PR's size makes review unreliable. Splitting it ensures each piece gets proper attention and nothing ships broken.
+
+---
+
+## Bug Summary
+
+| Severity | Count | Blocks Ship? |
+|----------|-------|-------------|
+| Critical (runtime crash) | 3 | Yes — C1, C2, C3 |
+| Critical (security) | 3 | Yes — C4, C5, C6 |
+| Major | 8 | Some — M1, M2, M3 are high risk |
+| Minor | 8 | No |
+| **Total** | **22** | |

docs/plans/PROJECT_STATUS.md

@@ -0,0 +1,462 @@
+# Voicebox Project Status & Roadmap
+
+> Last updated: 2026-03-12 | Current version: **v0.1.13** | 13.1k stars | 176 open issues | 28 open PRs
+
+---
+
+## Table of Contents
+
+1. [Architecture Overview](#architecture-overview)
+2. [Current State](#current-state)
+3. [Open PRs — Triage & Analysis](#open-prs--triage--analysis)
+4. [Open Issues — Categorized](#open-issues--categorized)
+5. [Existing Plan Documents — Status](#existing-plan-documents--status)
+6. [New Model Integration — Landscape](#new-model-integration--landscape)
+7. [Architectural Bottlenecks](#architectural-bottlenecks)
+8. [Recommended Priorities](#recommended-priorities)
+
+---
+
+## Architecture Overview
+
+```
+┌─────────────────────────────────────────────────────┐
+│  Tauri Shell (Rust)                                 │
+│  ┌───────────────────────────────────────────────┐  │
+│  │  React Frontend (app/)                        │  │
+│  │  Zustand stores · API client · Generation UI  │  │
+│  │  Stories Editor · Voice Profiles · Model Mgmt │  │
+│  └──────────────────────┬────────────────────────┘  │
+│                         │ HTTP :17493                │
+│  ┌──────────────────────▼────────────────────────┐  │
+│  │  FastAPI Backend (backend/)                   │  │
+│  │  ┌─────────────┐  ┌───────────┐  ┌─────────┐ │  │
+│  │  │ TTSBackend  │  │ STTBackend│  │ Profiles│ │  │
+│  │  │ (Protocol)  │  │ (Whisper) │  │ History │ │  │
+│  │  │  ┌────────┐ │  └───────────┘  │ Stories │ │  │
+│  │  │  │PyTorch │ │                  └─────────┘ │  │
+│  │  │  │or MLX  │ │                              │  │
+│  │  │  └────────┘ │                              │  │
+│  │  └─────────────┘                              │  │
+│  └───────────────────────────────────────────────┘  │
+└─────────────────────────────────────────────────────┘
+```
+
+### Key Files
+
+| Layer | File | Purpose |
+|-------|------|---------|
+| Backend entry | `backend/main.py` | FastAPI app, all API routes (~1700 lines) |
+| TTS protocol | `backend/backends/__init__.py:14-81` | `TTSBackend` Protocol definition |
+| TTS factory | `backend/backends/__init__.py:118-137` | Singleton backend selection (MLX vs PyTorch) |
+| PyTorch TTS | `backend/backends/pytorch_backend.py` | Qwen3-TTS via `qwen_tts` package |
+| MLX TTS | `backend/backends/mlx_backend.py` | Qwen3-TTS via `mlx_audio.tts` |
+| Platform detect | `backend/platform_detect.py` | Apple Silicon → MLX, else → PyTorch |
+| API types | `backend/models.py` | Pydantic request/response models |
+| Frontend API | `app/src/lib/api/client.ts` | Hand-written fetch wrapper |
+| Frontend types | `app/src/lib/api/types.ts` | TypeScript API types |
+| Generation form | `app/src/components/Generation/GenerationForm.tsx` | TTS generation UI |
+| Model manager | `app/src/components/ServerSettings/ModelManagement.tsx` | Model download/status UI |
+| Gen form hook | `app/src/lib/hooks/useGenerationForm.ts` | Form validation + submission |
+
+### How TTS Generation Works (Current Flow)
+
+```
+POST /generate
+  1. Look up voice profile from DB
+  2. Check model cache → if missing, trigger background download, return HTTP 202
+  3. Load model (lazy): tts_backend.load_model(model_size)
+  4. Create voice prompt: profiles.create_voice_prompt_for_profile()
+       → tts_backend.create_voice_prompt(audio_path, reference_text)
+  5. Generate: tts_backend.generate(text, voice_prompt, language, seed, instruct)
+  6. Save WAV → data/generations/{id}.wav
+  7. Insert history record in SQLite
+  8. Return GenerationResponse
+```
+
+---
+
+## Current State
+
+### What's Shipped (v0.1.13)
+
+- Qwen3-TTS voice cloning (1.7B and 0.6B models)
+- MLX backend for Apple Silicon, PyTorch for everything else
+- Voice profiles with multi-sample support
+- Stories editor (multi-track DAW timeline)
+- Whisper transcription (base, small, medium, large variants)
+- Model management UI with download progress (SSE)
+- Generation history with caching
+- Streaming generation endpoint (MLX only)
+- Delivery instructions (instruct parameter)
+
+### What's NOT Shipped But Has Code
+
+| Feature | Branch | Status |
+|---------|--------|--------|
+| External provider binaries (CUDA split) | `external-provider-binaries` | PR #33, significant work done, stale since Feb |
+| Dual server binaries | `feat/dual-server-binaries` | Branch exists, no PR |
+| Multi-sample fix | `fix-multi-sample` | Branch exists, no PR |
+| Model download notification fix | `fix-dl-notification-...` | Branch exists, no PR |
+
+### Hardcoded Qwen3-TTS Assumptions
+
+These are the specific coupling points that block multi-model support:
+
+| Location | What's Hardcoded |
+|----------|-----------------|
+| `backend/models.py:58` | `model_size` regex: `^(1\.7B\|0\.6B)$` |
+| `backend/main.py:611` | Default: `model_size or "1.7B"` |
+| `backend/main.py:1322-1365` | Model status list (2 Qwen + 4 Whisper) |
+| `backend/main.py:1523-1548` | Download trigger map |
+| `backend/main.py:1597-1628` | Delete map |
+| `backend/backends/pytorch_backend.py:65-68` | HF repo ID map |
+| `backend/backends/mlx_backend.py:41-44` | MLX repo ID map |
+| `backend/backends/__init__.py:118-137` | Single global TTS backend |
+| `app/src/lib/hooks/useGenerationForm.ts:17` | `modelSize: z.enum(['1.7B', '0.6B'])` |
+| `app/src/lib/hooks/useGenerationForm.ts:70-71` | `modelName = "qwen-tts-${data.modelSize}"` |
+| `app/src/components/Generation/GenerationForm.tsx:140-141` | Hardcoded "Qwen TTS" labels |
+| `app/src/components/ServerSettings/ModelManagement.tsx:166-213` | Filters by `qwen-tts` and `whisper` prefix |
+| `backend/utils/cache.py` | Voice prompt cache uses `torch.save()` |
+
+---
+
+## Open PRs — Triage & Analysis
+
+### Merge-Ready / Near-Ready (Bug Fixes & Small Features)
+
+| PR | Title | Risk | Notes |
+|----|-------|------|-------|
+| **#250** | docs: align local API port examples | None | Docs-only |
+| **#230** | docs: fix README grammar | None | Docs-only |
+| **#243** | a11y: screen reader and keyboard improvements | Low | Accessibility, no backend changes |
+| **#175** | Fix #134: duplicate profile name validation | Low | Simple validation |
+| **#178** | Fix #168 #140: generation error handling | Low | Error handling improvements |
+| **#152** | Fix: prevent crashes when HuggingFace unreachable | Medium | Monkey-patches HF hub; solves real offline bug (#150, #151) |
+| **#218** | fix: unify qwen tts cache dir on Windows | Low | Windows-specific path fix |
+| **#214** | fix: panic on launch from tokio::spawn | Low | Rust-side Tauri fix |
+| **#210** | fix: Linux NVIDIA GBM buffer crash | Low | Linux-specific, narrowly scoped |
+| **#88** | security: restrict CORS to known local origins | Low | Security hardening |
+
+### Significant Feature PRs
+
+| PR | Title | Complexity | Dependencies | Notes |
+|----|-------|-----------|--------------|-------|
+| **#97** | fix: pass language parameter to TTS models | Medium | None | **Critical bug** — language param was silently dropped. Adds `LANGUAGE_CODE_TO_NAME` mapping to both backends. Should be high priority. |
+| **#133** | feat: network access toggle | Low | None | Wires up existing plumbing (`--host 0.0.0.0`). Clean, small. |
+| **#238** | download cancel/clear UI + error panel | Medium | None | Adds cancel buttons, VS Code-style Problems panel, fixes whisper-large repo. Quality-of-life win. |
+| **#99** | feat: chunked TTS with quality selector | Medium | None | Solves the 500-char/2048-token limit. Sentence-aware splitting, crossfade concat, 44.1kHz upsampling. Addresses #191, #203, #69, #111. |
+| **#154** | feat: Audiobook tab | Medium | Depends on #99 concepts | Full audiobook workflow — chunked gen, preview, auto-save to Stories. New route + tab. |
+| **#91** | fix: CoreAudio device enumeration | Medium | None | macOS audio device handling. |
+
+### Architectural PRs (Need Careful Review)
+
+| PR | Title | Complexity | Notes |
+|----|-------|-----------|-------|
+| **#33** | CUDA GPU Support — External Provider Binaries | **Very High** | The big one. Splits monolithic backend into main app + downloadable provider executables (PyTorch CPU, CUDA). New provider management system, CI/CD for R2 uploads, provider settings UI. Created Feb 1, significant codebase. **This is the foundation for multi-model support** but is currently Qwen-only. |
+| **#225** | feat: custom HuggingFace model support | High | Adds `custom_models.py`, `custom:<slug>` model IDs, frontend model grouping (Built-in vs Custom). **Takes a different approach than #33** — keeps single backend but allows arbitrary HF repos. These two PRs may conflict architecturally. |
+| **#194** | feat: Hebrew + Chatterbox TTS | High | **First non-Qwen TTS model.** Adds `ChatterboxTTSBackend` alongside existing backends. Routes by language (`he` → Chatterbox, else → Qwen). Adds Hebrew Whisper models. Includes a lot of cleanup. Important precedent for multi-model. |
+| **#195** | feat: per-profile LoRA fine-tuning | **Very High** | Depends on #194. Training pipeline, adapter management, SSE progress, 15 new API endpoints. New DB tables. Forces PyTorch even on MLX systems for adapter inference. |
+| **#161** | feat: Docker + web deployment | High | 3-stage Dockerfile, SPA serving from FastAPI, docker-compose. Implements the Docker deployment plan. |
+| **#124** | Add Dockerfiles + docker-compose + docs | Medium | Earlier, simpler Docker attempt. Overlaps with #161. |
+| **#123** | added docker | Low | Minimal Docker PR. Overlaps with #161 and #124. |
+| **#227** | fix: harden input validation & file safety | Medium | Follow-up to #225. Atomic writes, threading locks, input validation. Good hardening but coupled to the custom models feature. |
+
+### PRs That Need Author Action / Are Stale
+
+| PR | Title | Notes |
+|----|-------|-------|
+| **#237** | fix: bundle qwen_tts source files in PyInstaller | Solves #212 but needs review for build system impact |
+| **#215** | Update prerequisites with Tauri deps | Branch is `main` — will have conflicts |
+| **#89** | Linux Support | Branch is `main` — will have conflicts. Broad scope. |
+| **#83** | Update download links for v0.1.12 | Outdated (we're on v0.1.13) |
+
+---
+
+## Open Issues — Categorized
+
+### GPU / Hardware Detection (19 issues)
+
+The single most reported category. Users on Windows with NVIDIA GPUs frequently report "GPU not detected."
+
+**Root causes (likely):**
+- PyInstaller binary doesn't bundle CUDA correctly → falls back to CPU
+- DirectML/Vulkan path not implemented (AMD on Windows)
+- Binary size limit means CUDA can't ship in the main release
+
+**Key issues:** #239, #222, #220, #217, #208, #198, #192, #167, #164, #141, #130, #127
+
+**Fix path:** PR #33 (external provider binaries) is designed to solve this. Ship a small main app, let users download the CUDA provider separately.
+
+### Model Downloads (20 issues)
+
+Second most reported. Users get stuck downloads, can't resume, no cancel button, no offline fallback.
+
+**Key issues:** #249, #240, #221, #216, #212, #181, #180, #159, #150, #149, #145, #143, #135, #134
+
+**Fix path:** PR #238 (cancel/clear UI), PR #152 (offline crash fix). Resume support not yet addressed.
+
+### Language Requests (18 issues)
+
+Strong demand for: Hindi (#245), Indonesian (#247), Dutch (#236), Hebrew (#199), Greek (#188), Portuguese (#183), Persian (#162), and many more.
+
+**Key issues:** #247, #245, #236, #211, #205, #199, #189, #188, #187, #183, #179, #162
+
+**Fix path:** PR #97 (pass language param — currently silently dropped!) is the prerequisite. Qwen3-TTS already supports many languages; the bug is that the language code isn't forwarded. Multi-model (#194 Chatterbox for Hebrew) expands coverage further.
+
+### New Model Requests (5 explicit issues)
+
+| Issue | Model Requested |
+|-------|----------------|
+| #226 | GGUF support |
+| #172 | VibeVoice |
+| #138 | Export to ONNX/Piper format |
+| #132 | LavaSR (transcription) |
+| #76 | (General model expansion) |
+
+Community is also vocally requesting: LuxTTS, Chatterbox, XTTS-v2, Fish Speech, CosyVoice, Kokoro on social media and in issue comments.
+
+### Long-Form / Chunking (5 issues)
+
+Users hitting the ~500 character practical limit.
+
+**Key issues:** #234 (queue system), #203 (500 char limit), #191 (auto-split), #111, #69
+
+**Fix path:** PR #99 (chunked TTS + quality selector) directly addresses this. PR #154 (Audiobook tab) builds on it.
+
+### Feature Requests (23 issues)
+
+Notable requests:
+- **#234** — Queue system for batch generation
+- **#182** — Concurrent/multi-thread generation
+- **#173** — Vocal intonation/inflection control
+- **#165** — Audiobook mode
+- **#144** — Copy text to clipboard
+- **#184** — Cancel button for progress bar
+- **#242** — Seed value pinning for consistency
+- **#228** — Always use 0.6B option
+- **#233** — Transcribe audio API improvements
+- **#235** — Finetuned Qwen3-TTS tokenizer
+
+### Bugs (19 issues)
+
+| Category | Issues |
+|----------|--------|
+| Generation failures | #248 (broken pipe), #219 (unsupported scalarType), #202 (clipping error), #170 (load failed) |
+| UI bugs | #231 (history not updating), #190 (mobile landing), #169 (blank interface) |
+| File operations | #207 (transcribe file error), #168 (no such file), #142 (download audio fail) |
+| Server lifecycle | #166 (server processes remain), #164 (no auto-update) |
+| Database | #174 (sqlite3 IntegrityError) |
+| Dependency | #131 (numpy ABI mismatch), #209 (import error) |
+
+---
+
+## Existing Plan Documents — Status
+
+| Document | Target Version | Status | Relevance |
+|----------|---------------|--------|-----------|
+| `TTS_PROVIDER_ARCHITECTURE.md` | v0.1.13 | **Partially implemented** in PR #33 | Core architecture for multi-model + CUDA distribution |
+| `EXTERNAL_PROVIDERS.md` | v0.2.0 | **Not started** | Remote server support. API path inconsistency with provider arch doc (`/v1/` vs `/tts/`) |
+| `MLX_AUDIO.md` | — | **Shipped** (the only one) | MLX backend is live. 0.6B MLX model still missing. |
+| `DOCKER_DEPLOYMENT.md` | v0.2.0 | **PR exists** (#161) | Waiting on review. No official images published. |
+| `OPENAI_SUPPORT.md` | v0.2.0 | **Not started** | OpenAI-compatible API layer. Linked to issue #10. Low complexity. |
+
+### Cross-Document Conflicts
+
+1. **API path inconsistency:** Provider arch uses `/tts/generate`, External providers uses `/v1/generate`, OpenAI compat uses `/v1/audio/speech`. Need to reconcile.
+2. **Docker vs. Provider split:** Docker doc assumes monolithic backend. Provider arch splits into separate binaries. Need to decide: does Docker run the monolith or individual providers?
+3. **Version targeting:** Provider arch targets v0.1.13 (current!) but isn't merged. Everything else targets v0.2.0.
+
+---
+
+## New Model Integration — Landscape
+
+### Models Worth Supporting (2026 SOTA)
+
+| Model | Cloning | Speed | Sample Rate | Languages | VRAM | Integration Ease | Repo |
+|-------|---------|-------|-------------|-----------|------|-----------------|------|
+| **LuxTTS** | 3s zero-shot | 150x RT, CPU ok | 48 kHz | English-first | <1 GB | Easy | `ysharma3501/LuxTTS` |
+| **Chatterbox** | 5s zero-shot | Sub-200ms streaming | 24-48 kHz | 23+ | Low | Medium | `resemble-ai/chatterbox` |
+| **XTTS-v2** | 6s zero-shot | Fast mid-GPU | 24 kHz | 17+ | Medium | Medium | `coqui/XTTS-v2` |
+| **Fish Speech** | 10-30s few-shot | Real-time | 24-44 kHz | 50+ | Medium | Medium | `fishaudio/fish-speech` |
+| **CosyVoice2-0.5B** | 3-10s zero-shot | Very fast | 24 kHz | Multilingual | Low | Easy | Alibaba HF org |
+| **Kokoro-82M** | 3s instant | CPU realtime | 24 kHz | English | Tiny | Medium | Kokoro repo |
+
+### What's Needed Architecturally for Multi-Model
+
+The current codebase assumes one TTS model family (Qwen3-TTS). Adding any new model requires:
+
+1. **Model type concept** — A `model_type` field (e.g. `qwen`, `luxtts`, `chatterbox`) alongside `model_size`. The `GenerationRequest` schema, frontend form, and all model config dicts need updating.
+
+2. **Multiple backend instances** — The singleton `get_tts_backend()` needs to become a registry. Different models have different voice prompt formats, different inference APIs, different sample rates.
+
+3. **Voice prompt format abstraction** — Qwen uses `torch.save()`-serialized tensors. LuxTTS uses `encode_prompt()` returning its own format. Chatterbox uses audio-path-based cloning. The cache system (`backend/utils/cache.py`) needs to handle heterogeneous formats.
+
+4. **Sample rate normalization** — Qwen outputs 24 kHz. LuxTTS outputs 48 kHz. The Stories editor and audio pipeline need to handle mixed rates.
+
+5. **Per-model capabilities** — Not all models support `instruct` (delivery instructions), not all support streaming, not all support the same languages. The UI needs to adapt.
+
+### PR #194 as Precedent
+
+The Hebrew/Chatterbox PR (#194) is the first attempt at multi-model. It takes a pragmatic approach: route by language (`he` → Chatterbox, else → Qwen). This works for one extra model but doesn't scale — what happens when you want Chatterbox for English too?
+
+### PR #225 as Alternative Approach
+
+The custom HuggingFace models PR (#225) takes a different angle: let users register arbitrary HF repos and attempt to load them through the existing Qwen backend. This is flexible but fragile — it assumes all models have the same API as Qwen3-TTS.
+
+### PR #33 as Foundation
+
+The external provider binaries PR (#33) has the most robust architecture for multi-model, since each provider is a separate process with its own dependencies. But it's complex, currently Qwen-only, and has been stale since early February.
+
+---
+
+## Architectural Bottlenecks
+
+### 1. Single Backend Singleton
+
+**File:** `backend/backends/__init__.py:118-137`
+
+The entire TTS system runs through one global `_tts_backend` instance. You literally cannot have two models loaded. This is the #1 blocker for multi-model support.
+
+### 2. `main.py` is 1700+ Lines
+
+All API routes, all model configs, all business logic in one file. Three separate hardcoded model config dicts that must stay in sync. Any multi-model change touches this file heavily.
+
+### 3. Model Config is Scattered
+
+Model identifiers, HF repo IDs, display names, and download logic are duplicated across:
+- `main.py` (3 separate dicts)
+- `pytorch_backend.py` (HF repo map)
+- `mlx_backend.py` (MLX repo map)
+- `GenerationForm.tsx` (UI labels)
+- `useGenerationForm.ts` (validation schema)
+- `ModelManagement.tsx` (prefix filters)
+
+There is no single source of truth for "what models does Voicebox support."
+
+### 4. Voice Prompt Cache Assumes PyTorch Tensors
+
+`backend/utils/cache.py` uses `torch.save()` / `torch.load()` for caching voice prompts. Models that don't use PyTorch tensors (LuxTTS, MLX-native models) can't use this cache.
+
+### 5. Frontend Assumes Qwen Model Sizes
+
+The generation form schema (`useGenerationForm.ts:17`) validates `model_size` as `'1.7B' | '0.6B'`. The model management UI filters by string prefix `qwen-tts`. Adding any model requires touching 3-4 frontend files.
+
+---
+
+## Recommended Priorities
+
+### Tier 1 — Ship Now (Bug Fixes & Critical Improvements)
+
+These PRs fix real user pain with low risk. Can be reviewed and merged quickly.
+
+| Priority | PR | Impact | Effort |
+|----------|-----|--------|--------|
+| 1 | **#97** — Pass language param to TTS | Fixes all non-English generation (18 language issues) | Low |
+| 2 | **#238** — Download cancel/clear UI | Addresses 20 download-related issues | Low |
+| 3 | **#152** — Offline mode crash fix | Fixes #150, #151 | Low |
+| 4 | **#99** — Chunked TTS + quality selector | Removes 500-char limit, addresses 5 issues | Medium |
+| 5 | **#218** — Windows HF cache dir fix | Windows-specific pain | Low |
+| 6 | **#175, #178** — Profile validation + error handling | Small fixes | Low |
+| 7 | **#250, #230** — Docs fixes | Zero risk | None |
+| 8 | **#133** — Network access toggle | Wires up existing code | Low |
+| 9 | **#88** — CORS restriction | Security improvement | Low |
+| 10 | **#214** — Tauri window close panic fix | Stability | Low |
+
+### Tier 2 — Next Release (v0.2.0 Foundations)
+
+These require more review but unlock major capabilities.
+
+| Priority | Item | Impact | Effort | Dependencies |
+|----------|------|--------|--------|-------------|
+| 1 | **PR #33** — External provider binaries | Solves GPU distribution (19 issues), foundation for multi-model | Very High | Needs rebase, thorough review |
+| 2 | **Multi-model abstraction layer** | Required before adding LuxTTS/Chatterbox/etc. | High | Informed by #33, #194, #225 |
+| 3 | **PR #161** — Docker deployment | Server/headless users | Medium | Independent of #33 |
+| 4 | **PR #194** — Hebrew + Chatterbox | First non-Qwen model, language expansion | High | Should align with multi-model abstraction |
+| 5 | **PR #154** — Audiobook tab | Significant feature for long-form users | Medium | Benefits from #99 (chunking) |
+
+### Tier 3 — Future (v0.3.0+)
+
+| Item | Notes |
+|------|-------|
+| LuxTTS integration | 48 kHz, low VRAM, but needs multi-model arch first |
+| XTTS-v2 / Fish Speech | Multilingual powerhouses |
+| OpenAI-compatible API (plan doc exists) | Low effort once API is stable |
+| LoRA fine-tuning (PR #195) | Complex, depends on #194 |
+| External/remote providers (plan doc exists) | Depends on provider architecture |
+| GGUF support (#226) | Depends on model ecosystem maturity |
+| Queue system (#234) | Batch generation |
+| Real-time streaming synthesis | MLX-only currently, needs PyTorch path |
+
+### Decision Point: Multi-Model Architecture
+
+Before adding any new TTS model, a decision is needed on *how*:
+
+**Option A — Provider Binary Split (PR #33 approach)**
+Each model family is a separate executable/process. Most isolated, most flexible, but most complex. Solves the CUDA distribution problem simultaneously.
+
+**Option B — In-Process Model Registry**
+Keep everything in one process but replace the singleton with a registry that can instantiate multiple `TTSBackend` implementations. Simpler, but doesn't solve binary size / CUDA distribution.
+
+**Option C — Hybrid (Recommended)**
+Use Option B for lightweight models (LuxTTS, Kokoro — small, CPU-friendly) that can coexist in-process. Use Option A for heavy models (CUDA Qwen3-TTS, Fish Speech) that need their own process/dependencies. The provider architecture from PR #33 becomes the escape hatch for heavy models, while light models are built-in.
+
+This matches how PR #194 already works (Chatterbox loaded in-process alongside Qwen) while keeping the door open for PR #33's provider split.
+
+---
+
+## Branch Inventory
+
+| Branch | PR | Status | Notes |
+|--------|-----|--------|-------|
+| `external-provider-binaries` | #33 | Open, stale | Major architecture work |
+| `feat/dual-server-binaries` | — | No PR | Related to provider split? |
+| `fix-multi-sample` | — | No PR | Voice profile multi-sample fix |
+| `fix-dl-notification-...` | — | No PR | Model download UX |
+| `improvements` | — | No PR | Unknown scope |
+| `stories` | — | No PR | Stories editor work? |
+| `windows-server-shutdown` | — | No PR | Windows lifecycle |
+| `model-dl-fix` | — | No PR | Model download fix |
+| `channels` | — | No PR | Audio channels |
+| `audio-export-entitlement-fix` | — | No PR | macOS entitlements |
+| `better-docs` | — | No PR | Documentation |
+
+---
+
+## Quick Reference: API Endpoints
+
+<details>
+<summary>All current endpoints (v0.1.13)</summary>
+
+| Endpoint | Method | Purpose |
+|----------|--------|---------|
+| `/health` | GET | Health check, model/GPU status |
+| `/profiles` | POST, GET | Create/list voice profiles |
+| `/profiles/{id}` | GET, PUT, DELETE | Profile CRUD |
+| `/profiles/{id}/samples` | POST, GET | Add/list voice samples |
+| `/profiles/{id}/avatar` | POST, GET, DELETE | Avatar management |
+| `/profiles/{id}/export` | GET | Export profile as ZIP |
+| `/profiles/import` | POST | Import profile from ZIP |
+| `/generate` | POST | Generate speech |
+| `/generate/stream` | POST | Stream speech (SSE) |
+| `/history` | GET | List generation history |
+| `/history/{id}` | GET, DELETE | Get/delete generation |
+| `/history/{id}/export` | GET | Export generation ZIP |
+| `/history/{id}/export-audio` | GET | Export audio only |
+| `/transcribe` | POST | Transcribe audio (Whisper) |
+| `/models/status` | GET | All model statuses |
+| `/models/download` | POST | Trigger model download |
+| `/models/{name}` | DELETE | Delete downloaded model |
+| `/models/load` | POST | Load model into memory |
+| `/models/unload` | POST | Unload model |
+| `/models/progress/{name}` | GET | SSE download progress |
+| `/tasks/active` | GET | Active downloads/generations |
+| `/stories` | POST, GET | Create/list stories |
+| `/stories/{id}` | GET, PUT, DELETE | Story CRUD |
+| `/stories/{id}/items` | POST, GET | Story items CRUD |
+| `/stories/{id}/export` | GET | Export story audio |
+| `/channels` | POST, GET | Audio channel CRUD |
+| `/channels/{id}` | PUT, DELETE | Channel update/delete |
+| `/cache/clear` | POST | Clear voice prompt cache |
+
+</details>

justfile

@@ -0,0 +1,189 @@
+# Voicebox development commands
+# Install: brew install just (or cargo install just)
+# Usage: just --list
+
+# Directories
+backend_dir := "backend"
+tauri_dir := "tauri"
+app_dir := "app"
+web_dir := "web"
+venv := backend_dir / "venv"
+venv_bin := venv / "bin"
+python := venv_bin / "python"
+pip := venv_bin / "pip"
+
+# Detect best python for venv creation
+system_python := `command -v python3.12 2>/dev/null || command -v python3.13 2>/dev/null || echo python3`
+
+# ─── Setup ────────────────────────────────────────────────────────────
+
+# Full project setup (python venv + JS deps + dev sidecar)
+setup: setup-python setup-js
+    @echo ""
+    @echo "Setup complete! Run: just dev"
+
+# Create venv and install Python dependencies
+setup-python:
+    #!/usr/bin/env bash
+    set -euo pipefail
+    if [ ! -d "{{ venv }}" ]; then
+        echo "Creating Python virtual environment..."
+        PY_MINOR=$({{ system_python }} -c "import sys; print(sys.version_info[1])")
+        if [ "$PY_MINOR" -gt 13 ]; then
+            echo "Warning: Python 3.$PY_MINOR detected. ML packages may not be compatible."
+            echo "Recommended: brew install [email protected]"
+        fi
+        {{ system_python }} -m venv {{ venv }}
+    fi
+    echo "Installing Python dependencies..."
+    {{ pip }} install --upgrade pip -q
+    {{ pip }} install -r {{ backend_dir }}/requirements.txt
+    # Apple Silicon: install MLX backend
+    if [ "$(uname -m)" = "arm64" ] && [ "$(uname)" = "Darwin" ]; then
+        echo "Detected Apple Silicon — installing MLX dependencies..."
+        {{ pip }} install -r {{ backend_dir }}/requirements-mlx.txt
+    fi
+    {{ pip }} install git+https://github.com/QwenLM/Qwen3-TTS.git
+    echo "Python environment ready."
+
+# Install JavaScript dependencies
+setup-js:
+    bun install
+
+# ─── Development ──────────────────────────────────────────────────────
+
+# Start backend + frontend for development (two processes, one terminal)
+dev: _ensure-venv _ensure-sidecar
+    #!/usr/bin/env bash
+    set -euo pipefail
+    trap 'kill 0' EXIT
+
+    echo "Starting backend on http://localhost:17493 ..."
+    {{ venv_bin }}/uvicorn backend.main:app --reload --port 17493 &
+    sleep 2
+
+    echo "Starting Tauri desktop app..."
+    cd {{ tauri_dir }} && bun run tauri dev &
+
+    wait
+
+# Start backend only
+dev-backend: _ensure-venv
+    {{ venv_bin }}/uvicorn backend.main:app --reload --port 17493
+
+# Start Tauri desktop app only (backend must be running separately)
+dev-frontend: _ensure-sidecar
+    cd {{ tauri_dir }} && bun run tauri dev
+
+# Start backend + web app (no Tauri)
+dev-web: _ensure-venv
+    #!/usr/bin/env bash
+    set -euo pipefail
+    trap 'kill 0' EXIT
+    {{ venv_bin }}/uvicorn backend.main:app --reload --port 17493 &
+    sleep 2
+    cd {{ web_dir }} && bun run dev &
+    wait
+
+# Kill all dev processes
+kill:
+    -pkill -f "uvicorn backend.main:app" 2>/dev/null || true
+    -pkill -f "vite" 2>/dev/null || true
+    @echo "Dev processes killed."
+
+# ─── Build ────────────────────────────────────────────────────────────
+
+# Build everything (server binary + desktop app)
+build: build-server build-tauri
+
+# Build Python server binary
+build-server: _ensure-venv
+    PATH="{{ venv_bin }}:$PATH" ./scripts/build-server.sh
+
+# Build Tauri desktop app
+build-tauri:
+    cd {{ tauri_dir }} && bun run tauri build
+
+# Build web app
+build-web:
+    cd {{ web_dir }} && bun run build
+
+# ─── Code Quality ────────────────────────────────────────────────────
+
+# Run all checks (lint + format + typecheck)
+check:
+    bun run check
+
+# Lint with Biome
+lint:
+    bun run lint
+
+# Format with Biome
+format:
+    bun run format
+
+# Fix lint + format issues
+fix:
+    bun run check:fix
+
+# ─── Database ─────────────────────────────────────────────────────────
+
+# Initialize SQLite database
+db-init: _ensure-venv
+    cd {{ backend_dir }} && {{ python }} -c "from database import init_db; init_db()"
+
+# Reset database (delete + reinit)
+db-reset:
+    rm -f {{ backend_dir }}/data/voicebox.db
+    just db-init
+
+# ─── Utilities ────────────────────────────────────────────────────────
+
+# Generate TypeScript API client (backend must be running)
+generate-api:
+    ./scripts/generate-api.sh
+
+# Open API docs in browser
+docs:
+    open http://localhost:17493/docs 2>/dev/null || xdg-open http://localhost:17493/docs
+
+# Tail backend logs
+logs:
+    tail -f {{ backend_dir }}/logs/*.log 2>/dev/null || echo "No log files found"
+
+# ─── Clean ────────────────────────────────────────────────────────────
+
+# Clean build artifacts
+clean:
+    rm -rf {{ tauri_dir }}/src-tauri/target/release
+    rm -rf {{ web_dir }}/dist
+    rm -rf {{ app_dir }}/dist
+
+# Clean Python venv and cache
+clean-python:
+    rm -rf {{ venv }}
+    find {{ backend_dir }} -type d -name "__pycache__" -exec rm -rf {} + 2>/dev/null || true
+
+# Nuclear clean (everything including node_modules)
+clean-all: clean clean-python
+    rm -rf node_modules
+    rm -rf {{ app_dir }}/node_modules
+    rm -rf {{ tauri_dir }}/node_modules
+    rm -rf {{ web_dir }}/node_modules
+    cd {{ tauri_dir }}/src-tauri && cargo clean
+
+# ─── Internal ─────────────────────────────────────────────────────────
+
+# Ensure venv exists (prompt to run setup if not)
+[private]
+_ensure-venv:
+    #!/usr/bin/env bash
+    if [ ! -d "{{ venv }}" ]; then
+        echo "Python venv not found. Run: just setup"
+        exit 1
+    fi
+
+# Ensure Tauri dev sidecar placeholder exists
+[private]
+_ensure-sidecar:
+    bun run setup:dev

landing/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/landing",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "description": "Landing page for voicebox.sh",
   "scripts": {
     "dev": "bun --bun next dev --turbo",

landing/src/app/layout.tsx

@@ -1,6 +1,7 @@
 import type { Metadata } from 'next';
 import { Inter } from 'next/font/google';
 import './globals.css';
+import { Banner } from '@/components/Banner';
 import { Footer } from '@/components/Footer';
 import { Header } from '@/components/Header';
 
@@ -31,6 +32,7 @@ export default function RootLayout({ children }: { children: React.ReactNode })
     <html lang="en" suppressHydrationWarning className="dark">
       <body className={inter.variable}>
         <div className="relative min-h-screen bg-background font-sans flex flex-col">
+          <Banner />
           <Header />
           <main className="container mx-auto px-4 sm:px-6 md:px-4 flex-1 py-4 sm:py-6 md:py-0">
             {children}

landing/src/components/Banner.tsx

@@ -0,0 +1,25 @@
+import { ArrowRight } from 'lucide-react';
+
+export function Banner() {
+  return (
+    <div className="bg-primary/[0.06] border-b border-border backdrop-blur-sm">
+      <div className="container mx-auto px-4">
+        <div className="flex items-center justify-center h-10 text-sm">
+          <a
+            href="https://spacebot.sh"
+            target="_blank"
+            rel="noopener noreferrer"
+            className="flex items-center gap-2 text-muted-foreground hover:text-foreground transition-colors group"
+          >
+            <span>
+              Also by the creator of Voicebox:{' '}
+              <strong className="text-foreground/90">Spacebot</strong>, an AI agent OS for teams.
+              Connect Discord, Slack, or Telegram in one click.
+            </span>
+            <ArrowRight className="h-3.5 w-3.5 transition-transform group-hover:translate-x-0.5" />
+          </a>
+        </div>
+      </div>
+    </div>
+  );
+}

package.json

@@ -1,6 +1,6 @@
 {
   "name": "voicebox",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "private": true,
   "workspaces": [
     "app",

requirements.txt

@@ -0,0 +1,9 @@
+uvicorn
+fastapi
+sqlalchemy
+torch
+torchvision
+soundfile
+librosa
+python-multipart
+huggingface_hub

scripts/setup-dev-sidecar.js

@@ -1,4 +1,5 @@
 #!/usr/bin/env node
+
 /**
  * Creates placeholder sidecar binaries for development mode.
  *
@@ -9,10 +10,10 @@
  * The actual server should be started separately with `bun run dev:server`.
  */
 
-import { existsSync, mkdirSync, writeFileSync, statSync } from 'fs';
-import { join, dirname } from 'path';
-import { fileURLToPath } from 'url';
 import { execSync } from 'child_process';
+import { existsSync, mkdirSync, statSync, writeFileSync } from 'fs';
+import { dirname, join } from 'path';
+import { fileURLToPath } from 'url';
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
@@ -55,7 +56,9 @@ function createPlaceholderBinary(targetTriple) {
     try {
       const stats = statSync(binaryPath);
       if (stats.size > MIN_REAL_BINARY_SIZE) {
-        console.log(`Real binary already exists: ${binaryName} (${(stats.size / 1024 / 1024).toFixed(1)} MB)`);
+        console.log(
+          `Real binary already exists: ${binaryName} (${(stats.size / 1024 / 1024).toFixed(1)} MB)`,
+        );
         return;
       }
     } catch {
@@ -73,52 +76,275 @@ function createPlaceholderBinary(targetTriple) {
     // This is the smallest valid PE that Windows will accept
     const minimalPE = Buffer.from([
       // DOS Header
-      0x4D, 0x5A, 0x90, 0x00, 0x03, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xFF, 0xFF, 0x00, 0x00,
-      0xB8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-      0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-      0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x00, 0x00, 0x00,
+      0x4d,
+      0x5a,
+      0x90,
+      0x00,
+      0x03,
+      0x00,
+      0x00,
+      0x00,
+      0x04,
+      0x00,
+      0x00,
+      0x00,
+      0xff,
+      0xff,
+      0x00,
+      0x00,
+      0xb8,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x40,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x80,
+      0x00,
+      0x00,
+      0x00,
       // DOS Stub
-      0x0E, 0x1F, 0xBA, 0x0E, 0x00, 0xB4, 0x09, 0xCD, 0x21, 0xB8, 0x01, 0x4C, 0xCD, 0x21, 0x54, 0x68,
-      0x69, 0x73, 0x20, 0x70, 0x72, 0x6F, 0x67, 0x72, 0x61, 0x6D, 0x20, 0x63, 0x61, 0x6E, 0x6E, 0x6F,
-      0x74, 0x20, 0x62, 0x65, 0x20, 0x72, 0x75, 0x6E, 0x20, 0x69, 0x6E, 0x20, 0x44, 0x4F, 0x53, 0x20,
-      0x6D, 0x6F, 0x64, 0x65, 0x2E, 0x0D, 0x0D, 0x0A, 0x24, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+      0x0e,
+      0x1f,
+      0xba,
+      0x0e,
+      0x00,
+      0xb4,
+      0x09,
+      0xcd,
+      0x21,
+      0xb8,
+      0x01,
+      0x4c,
+      0xcd,
+      0x21,
+      0x54,
+      0x68,
+      0x69,
+      0x73,
+      0x20,
+      0x70,
+      0x72,
+      0x6f,
+      0x67,
+      0x72,
+      0x61,
+      0x6d,
+      0x20,
+      0x63,
+      0x61,
+      0x6e,
+      0x6e,
+      0x6f,
+      0x74,
+      0x20,
+      0x62,
+      0x65,
+      0x20,
+      0x72,
+      0x75,
+      0x6e,
+      0x20,
+      0x69,
+      0x6e,
+      0x20,
+      0x44,
+      0x4f,
+      0x53,
+      0x20,
+      0x6d,
+      0x6f,
+      0x64,
+      0x65,
+      0x2e,
+      0x0d,
+      0x0d,
+      0x0a,
+      0x24,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
       // PE Signature
-      0x50, 0x45, 0x00, 0x00,
+      0x50,
+      0x45,
+      0x00,
+      0x00,
       // COFF Header (x64)
-      0x64, 0x86, // Machine: AMD64
-      0x01, 0x00, // NumberOfSections: 1
-      0x00, 0x00, 0x00, 0x00, // TimeDateStamp
-      0x00, 0x00, 0x00, 0x00, // PointerToSymbolTable
-      0x00, 0x00, 0x00, 0x00, // NumberOfSymbols
-      0xF0, 0x00, // SizeOfOptionalHeader
-      0x22, 0x00, // Characteristics: EXECUTABLE_IMAGE | LARGE_ADDRESS_AWARE
+      0x64,
+      0x86, // Machine: AMD64
+      0x01,
+      0x00, // NumberOfSections: 1
+      0x00,
+      0x00,
+      0x00,
+      0x00, // TimeDateStamp
+      0x00,
+      0x00,
+      0x00,
+      0x00, // PointerToSymbolTable
+      0x00,
+      0x00,
+      0x00,
+      0x00, // NumberOfSymbols
+      0xf0,
+      0x00, // SizeOfOptionalHeader
+      0x22,
+      0x00, // Characteristics: EXECUTABLE_IMAGE | LARGE_ADDRESS_AWARE
       // Optional Header (PE32+)
-      0x0B, 0x02, // Magic: PE32+
-      0x00, 0x00, // Linker version
-      0x00, 0x00, 0x00, 0x00, // SizeOfCode
-      0x00, 0x00, 0x00, 0x00, // SizeOfInitializedData
-      0x00, 0x00, 0x00, 0x00, // SizeOfUninitializedData
-      0x00, 0x10, 0x00, 0x00, // AddressOfEntryPoint
-      0x00, 0x00, 0x00, 0x00, // BaseOfCode
-      0x00, 0x00, 0x00, 0x40, 0x01, 0x00, 0x00, 0x00, // ImageBase
-      0x00, 0x10, 0x00, 0x00, // SectionAlignment
-      0x00, 0x02, 0x00, 0x00, // FileAlignment
-      0x06, 0x00, 0x00, 0x00, // OS version
-      0x00, 0x00, 0x00, 0x00, // Image version
-      0x06, 0x00, 0x00, 0x00, // Subsystem version
-      0x00, 0x00, 0x00, 0x00, // Win32VersionValue
-      0x00, 0x20, 0x00, 0x00, // SizeOfImage
-      0x00, 0x02, 0x00, 0x00, // SizeOfHeaders
-      0x00, 0x00, 0x00, 0x00, // CheckSum
-      0x03, 0x00, // Subsystem: CONSOLE
-      0x60, 0x01, // DllCharacteristics
+      0x0b,
+      0x02, // Magic: PE32+
+      0x00,
+      0x00, // Linker version
+      0x00,
+      0x00,
+      0x00,
+      0x00, // SizeOfCode
+      0x00,
+      0x00,
+      0x00,
+      0x00, // SizeOfInitializedData
+      0x00,
+      0x00,
+      0x00,
+      0x00, // SizeOfUninitializedData
+      0x00,
+      0x10,
+      0x00,
+      0x00, // AddressOfEntryPoint
+      0x00,
+      0x00,
+      0x00,
+      0x00, // BaseOfCode
+      0x00,
+      0x00,
+      0x00,
+      0x40,
+      0x01,
+      0x00,
+      0x00,
+      0x00, // ImageBase
+      0x00,
+      0x10,
+      0x00,
+      0x00, // SectionAlignment
+      0x00,
+      0x02,
+      0x00,
+      0x00, // FileAlignment
+      0x06,
+      0x00,
+      0x00,
+      0x00, // OS version
+      0x00,
+      0x00,
+      0x00,
+      0x00, // Image version
+      0x06,
+      0x00,
+      0x00,
+      0x00, // Subsystem version
+      0x00,
+      0x00,
+      0x00,
+      0x00, // Win32VersionValue
+      0x00,
+      0x20,
+      0x00,
+      0x00, // SizeOfImage
+      0x00,
+      0x02,
+      0x00,
+      0x00, // SizeOfHeaders
+      0x00,
+      0x00,
+      0x00,
+      0x00, // CheckSum
+      0x03,
+      0x00, // Subsystem: CONSOLE
+      0x60,
+      0x01, // DllCharacteristics
       // Stack/Heap sizes (8 bytes each for PE32+)
-      0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00,
-      0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-      0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00,
-      0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-      0x00, 0x00, 0x00, 0x00, // LoaderFlags
-      0x10, 0x00, 0x00, 0x00, // NumberOfRvaAndSizes
+      0x00,
+      0x00,
+      0x10,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x10,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00,
+      0x00, // LoaderFlags
+      0x10,
+      0x00,
+      0x00,
+      0x00, // NumberOfRvaAndSizes
     ]);
 
     // Pad to 512 bytes minimum for valid PE
@@ -138,19 +364,8 @@ exit 1
 }
 
 function main() {
-  console.log('Setting up development sidecar...');
-  console.log('');
-
   const targetTriple = getTargetTriple();
-  console.log(`Platform: ${targetTriple}`);
-
   createPlaceholderBinary(targetTriple);
-
-  console.log('');
-  console.log('Sidecar setup complete.');
-  console.log('For development, start the Python server in a separate terminal:');
-  console.log('  bun run dev:server');
-  console.log('');
 }
 
 main();

scripts/split_binary.py

@@ -0,0 +1,82 @@
+"""
+Split a large binary into chunks for GitHub Releases (<2 GB each).
+
+Usage:
+    python scripts/split_binary.py backend/dist/voicebox-server-cuda.exe
+    python scripts/split_binary.py backend/dist/voicebox-server-cuda.exe --chunk-size 1900000000
+    python scripts/split_binary.py backend/dist/voicebox-server-cuda.exe --output release-assets/
+
+The script produces:
+    - voicebox-server-cuda.part00.exe, .part01.exe, ...  (binary chunks)
+    - voicebox-server-cuda.sha256      (SHA-256 checksum of the complete file)
+    - voicebox-server-cuda.manifest    (ordered list of part filenames)
+"""
+
+import argparse
+import hashlib
+import sys
+from pathlib import Path
+
+
+def split(input_path: Path, chunk_size: int, output_dir: Path):
+    output_dir.mkdir(parents=True, exist_ok=True)
+    data = input_path.read_bytes()
+    total_size = len(data)
+
+    # Write SHA-256 of the complete file
+    sha256 = hashlib.sha256(data).hexdigest()
+    checksum_file = output_dir / f"{input_path.stem}.sha256"
+    checksum_file.write_text(f"{sha256}  {input_path.name}\n")
+
+    # Split into chunks
+    parts = []
+    for i in range(0, total_size, chunk_size):
+        part_index = len(parts)
+        part_name = f"{input_path.stem}.part{part_index:02d}{input_path.suffix}"
+        part_path = output_dir / part_name
+        part_path.write_bytes(data[i:i + chunk_size])
+        parts.append(part_name)
+
+    # Write manifest (ordered list of part filenames)
+    manifest_file = output_dir / f"{input_path.stem}.manifest"
+    manifest_file.write_text("\n".join(parts) + "\n")
+
+    print(f"Input:    {input_path} ({total_size / (1024**3):.2f} GB)")
+    print(f"Output:   {output_dir}/")
+    print(f"Parts:    {len(parts)} (chunk size: {chunk_size / (1024**3):.2f} GB)")
+    print(f"SHA-256:  {sha256}")
+    print(f"Manifest: {manifest_file.name}")
+    for p in parts:
+        size = (output_dir / p).stat().st_size
+        print(f"  {p}  ({size / (1024**3):.2f} GB)")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Split a large binary into chunks for GitHub Releases"
+    )
+    parser.add_argument("input", type=Path, help="Path to the binary file to split")
+    parser.add_argument(
+        "--chunk-size",
+        type=int,
+        default=1_900_000_000,  # 1.9 GB — safely under 2 GB GitHub limit
+        help="Maximum chunk size in bytes (default: 1.9 GB)",
+    )
+    parser.add_argument(
+        "--output",
+        type=Path,
+        default=None,
+        help="Output directory (default: same directory as input)",
+    )
+    args = parser.parse_args()
+
+    if not args.input.exists():
+        print(f"Error: {args.input} does not exist", file=sys.stderr)
+        sys.exit(1)
+
+    output_dir = args.output or args.input.parent
+    split(args.input, args.chunk_size, output_dir)
+
+
+if __name__ == "__main__":
+    main()

tauri/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/tauri",
   "private": true,
-  "version": "0.1.12",
+  "version": "0.1.13",
   "type": "module",
   "scripts": {
     "dev": "vite",

tauri/src-tauri/Cargo.lock

@@ -5041,7 +5041,7 @@ checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
 
 [[package]]
 name = "voicebox"
-version = "0.1.11"
+version = "0.1.12"
 dependencies = [
  "base64 0.22.1",
  "core-foundation-sys",

tauri/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "voicebox"
-version = "0.1.12"
+version = "0.1.13"
 description = "A production-quality desktop app for Qwen3-TTS voice cloning and generation"
 authors = ["you"]
 license = ""

tauri/src-tauri/src/main.rs

@@ -178,6 +178,56 @@ async fn start_server(
     println!("Data directory: {:?}", data_dir);
     println!("Remote mode: {}", remote.unwrap_or(false));
 
+    // Check for CUDA backend binary in data directory
+    let cuda_binary = {
+        let backends_dir = data_dir.join("backends");
+        let cuda_name = if cfg!(windows) {
+            "voicebox-server-cuda.exe"
+        } else {
+            "voicebox-server-cuda"
+        };
+        let path = backends_dir.join(cuda_name);
+        if path.exists() {
+            println!("Found CUDA backend binary at {:?}", path);
+
+            // Version check: run --version and compare to app version
+            let app_version = app.config().version.clone().unwrap_or_default();
+            let version_ok = match std::process::Command::new(&path)
+                .arg("--version")
+                .output()
+            {
+                Ok(output) => {
+                    // Output format: "voicebox-server X.Y.Z\n"
+                    let version_str = String::from_utf8_lossy(&output.stdout);
+                    let binary_version = version_str.trim().split_whitespace().last().unwrap_or("");
+                    if binary_version == app_version {
+                        println!("CUDA binary version {} matches app version", binary_version);
+                        true
+                    } else {
+                        println!(
+                            "CUDA binary version mismatch: binary={}, app={}. Falling back to CPU.",
+                            binary_version, app_version
+                        );
+                        false
+                    }
+                }
+                Err(e) => {
+                    println!("Failed to check CUDA binary version: {}. Falling back to CPU.", e);
+                    false
+                }
+            };
+
+            if version_ok {
+                Some(path)
+            } else {
+                None
+            }
+        } else {
+            println!("No CUDA backend found, using bundled CPU binary");
+            None
+        }
+    };
+
     let sidecar_result = app.shell().sidecar("voicebox-server");
 
     let mut sidecar = match sidecar_result {
@@ -216,22 +266,32 @@ async fn start_server(
 
     println!("Sidecar command created successfully");
 
-    // Pass data directory and port to Python server
-    sidecar = sidecar.args([
-        "--data-dir",
-        data_dir
-            .to_str()
-            .ok_or_else(|| "Invalid data dir path".to_string())?,
-        "--port",
-        &SERVER_PORT.to_string(),
-    ]);
+    // Build common args
+    let data_dir_str = data_dir
+        .to_str()
+        .ok_or_else(|| "Invalid data dir path".to_string())?
+        .to_string();
+    let port_str = SERVER_PORT.to_string();
+    let is_remote = remote.unwrap_or(false);
 
-    if remote.unwrap_or(false) {
-        sidecar = sidecar.args(["--host", "0.0.0.0"]);
-    }
-
-    println!("Spawning server process...");
-    let spawn_result = sidecar.spawn();
+    // If CUDA binary exists, launch it directly instead of the bundled sidecar
+    let spawn_result = if let Some(ref cuda_path) = cuda_binary {
+        println!("Launching CUDA backend: {:?}", cuda_path);
+        let mut cmd = app.shell().command(cuda_path.to_str().unwrap());
+        cmd = cmd.args(["--data-dir", &data_dir_str, "--port", &port_str]);
+        if is_remote {
+            cmd = cmd.args(["--host", "0.0.0.0"]);
+        }
+        cmd.spawn()
+    } else {
+        // Use the bundled CPU sidecar
+        sidecar = sidecar.args(["--data-dir", &data_dir_str, "--port", &port_str]);
+        if is_remote {
+            sidecar = sidecar.args(["--host", "0.0.0.0"]);
+        }
+        println!("Spawning server process...");
+        sidecar.spawn()
+    };
 
     let (mut rx, child) = match spawn_result {
         Ok(result) => result,
@@ -549,6 +609,25 @@ async fn stop_server(state: State<'_, ServerState>) -> Result<(), String> {
     Ok(())
 }
 
+#[command]
+async fn restart_server(
+    app: tauri::AppHandle,
+    state: State<'_, ServerState>,
+) -> Result<String, String> {
+    println!("restart_server: stopping current server...");
+
+    // Stop the current server
+    stop_server(state.clone()).await?;
+
+    // Wait for port to be released
+    println!("restart_server: waiting for port release...");
+    tokio::time::sleep(tokio::time::Duration::from_millis(1000)).await;
+
+    // Start server again (will auto-detect CUDA binary)
+    println!("restart_server: starting server...");
+    start_server(app, state, None).await
+}
+
 #[command]
 fn set_keep_server_running(state: State<'_, ServerState>, keep_running: bool) {
     *state.keep_running_on_close.lock().unwrap() = keep_running;
@@ -640,6 +719,7 @@ pub fn run() {
         .invoke_handler(tauri::generate_handler![
             start_server,
             stop_server,
+            restart_server,
             set_keep_server_running,
             start_system_audio_capture,
             stop_system_audio_capture,

tauri/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Voicebox",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "identifier": "sh.voicebox.app",
   "build": {
     "beforeDevCommand": "bun run dev",
@@ -12,7 +12,7 @@
   "bundle": {
     "active": true,
     "targets": "all",
-    "createUpdaterArtifacts": true,
+    "createUpdaterArtifacts": false,
     "externalBin": ["binaries/voicebox-server"],
     "icon": [
       "icons/32x32.png",

tauri/src/platform/filesystem.ts

@@ -2,40 +2,25 @@ import type { PlatformFilesystem, FileFilter } from '@/platform/types';
 
 export const tauriFilesystem: PlatformFilesystem = {
   async saveFile(filename: string, blob: Blob, filters?: FileFilter[]) {
-    try {
-      const { save } = await import('@tauri-apps/plugin-dialog');
-      const filePath = await save({
-        defaultPath: filename,
-        filters: filters || [],
-      });
+    const { save } = await import('@tauri-apps/plugin-dialog');
+    const { writeFile } = await import('@tauri-apps/plugin-fs');
 
-      if (filePath) {
-        let resolvedPath = '';
-        if (typeof filePath === 'string') {
-          resolvedPath = filePath;
-        } else if (filePath && typeof filePath === 'object' && 'path' in filePath) {
-          resolvedPath = (filePath as { path: string }).path;
-        }
+    const filePath = await save({
+      defaultPath: filename,
+      filters: filters || [],
+    });
 
-        if (!resolvedPath) {
-          throw new Error('Failed to resolve save path');
-        }
+    if (!filePath) return; // User cancelled the dialog
 
-        const { writeFile } = await import('@tauri-apps/plugin-fs');
-        const arrayBuffer = await blob.arrayBuffer();
-        await writeFile(resolvedPath, new Uint8Array(arrayBuffer));
-      }
-    } catch (error) {
-      console.error('Failed to use Tauri dialog, falling back to browser download:', error);
-      // Fall back to browser download if Tauri dialog fails
-      const url = window.URL.createObjectURL(blob);
-      const a = document.createElement('a');
-      a.href = url;
-      a.download = filename;
-      document.body.appendChild(a);
-      a.click();
-      window.URL.revokeObjectURL(url);
-      document.body.removeChild(a);
+    const resolvedPath = typeof filePath === 'string'
+      ? filePath
+      : (filePath as { path: string }).path;
+
+    if (!resolvedPath) {
+      throw new Error('Failed to resolve save path from dialog');
     }
+
+    const arrayBuffer = await blob.arrayBuffer();
+    await writeFile(resolvedPath, new Uint8Array(arrayBuffer));
   },
 };

tauri/src/platform/lifecycle.ts

@@ -1,5 +1,5 @@
 import { invoke } from '@tauri-apps/api/core';
-import { listen, emit } from '@tauri-apps/api/event';
+import { emit, listen } from '@tauri-apps/api/event';
 import type { PlatformLifecycle } from '@/platform/types';
 
 class TauriLifecycle implements PlatformLifecycle {
@@ -27,6 +27,18 @@ class TauriLifecycle implements PlatformLifecycle {
     }
   }
 
+  async restartServer(): Promise<string> {
+    try {
+      const result = await invoke<string>('restart_server');
+      console.log('Server restarted:', result);
+      this.onServerReady?.();
+      return result;
+    } catch (error) {
+      console.error('Failed to restart server:', error);
+      throw error;
+    }
+  }
+
   async setKeepServerRunning(keepRunning: boolean): Promise<void> {
     try {
       await invoke('set_keep_server_running', { keepRunning });

web/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/web",
   "private": true,
-  "version": "0.1.12",
+  "version": "0.1.13",
   "type": "module",
   "scripts": {
     "dev": "vite",
@@ -21,6 +21,7 @@
     "@types/react-dom": "^18.3.0",
     "@typescript-eslint/eslint-plugin": "^7.0.0",
     "@typescript-eslint/parser": "^7.0.0",
+    "@tailwindcss/vite": "^4.0.0",
     "@vitejs/plugin-react": "^4.3.0",
     "eslint": "^8.57.0",
     "eslint-plugin-react-hooks": "^4.6.0",

web/src/platform/lifecycle.ts

@@ -15,6 +15,11 @@ class WebLifecycle implements PlatformLifecycle {
     // No-op for web - server is managed externally
   }
 
+  async restartServer(): Promise<string> {
+    // No-op for web - server is managed externally
+    return import.meta.env.VITE_SERVER_URL || 'http://localhost:17493';
+  }
+
   async setKeepServerRunning(_keep: boolean): Promise<void> {
     // No-op for web
   }

web/vite.config.ts

@@ -1,9 +1,10 @@
 import path from 'node:path';
 import react from '@vitejs/plugin-react';
+import tailwindcss from '@tailwindcss/vite';
 import { defineConfig } from 'vite';
 
 export default defineConfig({
-  plugins: [react()],
+  plugins: [react(), tailwindcss()],
   resolve: {
     alias: {
       '@': path.resolve(__dirname, '../app/src'),