fix: pin numba for release CI wheel compatibility

Bump version: 0.1.12 → 0.1.13
docs: point download links to latest release
2026-09-28 14:45:16 -07:00 · 2026-02-23 12:18:34 -08:00 · 2026-02-23 11:24:46 -08:00 · 2026-02-23 11:23:30 -08:00 · 2026-02-23 11:13:12 -08:00 · 2026-02-23 11:12:34 -08:00
51 changed files with 1530 additions and 1954 deletions
@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.1.12
+current_version = 0.1.13
 commit = True
 tag = True
 tag_name = v{new_version}
@@ -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
@@ -4,7 +4,7 @@ on:
  workflow_dispatch:
  push:
    tags:
-      - 'v*'
+      - "v*"

 jobs:
  release:
@@ -14,22 +14,22 @@ jobs:
      fail-fast: false
      matrix:
        include:
-          - platform: 'macos-latest'
-            args: '--target aarch64-apple-darwin'
-            python-version: '3.12'
-            backend: 'mlx'
-          - platform: 'macos-15-intel'
-            args: '--target x86_64-apple-darwin'
-            python-version: '3.12'
-            backend: 'pytorch'
+          - platform: "macos-latest"
+            args: "--target aarch64-apple-darwin"
+            python-version: "3.12"
+            backend: "mlx"
+          - platform: "macos-15-intel"
+            args: "--target x86_64-apple-darwin"
+            python-version: "3.12"
+            backend: "pytorch"
          # - platform: 'ubuntu-22.04'
          #   args: ''
          #   python-version: '3.12'
          #   backend: 'pytorch'
-          - platform: 'windows-latest'
-            args: ''
-            python-version: '3.12'
-            backend: 'pytorch'
+          - platform: "windows-latest"
+            args: ""
+            python-version: "3.12"
+            backend: "pytorch"

    runs-on: ${{ matrix.platform }}

@@ -53,7 +53,7 @@ jobs:
        uses: actions/setup-python@v5
        with:
          python-version: ${{ matrix.python-version }}
-          cache: 'pip'
+          cache: "pip"

      - name: Install Python dependencies
        run: |
@@ -66,24 +66,24 @@ jobs:
        run: |
          pip install -r backend/requirements-mlx.txt

+      # - name: Install PyTorch with CUDA (Windows only)
+      #   if: matrix.platform == 'windows-latest'
+      #   run: |
+      #     pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-deps
+      #     pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
+
      - name: Build Python server (Linux/macOS)
        if: matrix.platform != 'windows-latest'
        run: |
          chmod +x scripts/build-server.sh
          ./scripts/build-server.sh

-      - name: Build CPU Python server (Windows)
+      - name: Build Python server (Windows)
        if: matrix.platform == 'windows-latest'
        shell: bash
        run: |
          cd backend
-
-          echo "Installing CPU-only PyTorch..."
-          pip uninstall -y torch torchvision torchaudio
-          pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-
-          echo "Building CPU server binary..."
-          python build_binary.py cpu
+          python build_binary.py

          # Get platform tuple
          PLATFORM=$(rustc --print host-tuple)
@@ -91,31 +91,9 @@ jobs:
          # Create binaries directory
          mkdir -p ../tauri/src-tauri/binaries

-          # Copy CPU version (default for installer)
+          # Copy with platform suffix
          cp dist/voicebox-server.exe ../tauri/src-tauri/binaries/voicebox-server-${PLATFORM}.exe
-          echo "Built CPU server: voicebox-server-${PLATFORM}.exe (~500MB)"
-
-      - name: Build CUDA Python server (Windows)
-        if: matrix.platform == 'windows-latest'
-        shell: bash
-        run: |
-          cd backend
-
-          echo "Installing CUDA PyTorch..."
-          pip uninstall -y torch torchvision torchaudio
-          pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-deps
-          pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
-
-          echo "Building CUDA server binary..."
-          python build_binary.py cuda
-
-          # Get platform tuple
-          PLATFORM=$(rustc --print host-tuple)
-
-          # Copy CUDA version for separate upload
-          mkdir -p cuda-release
-          cp dist/voicebox-server-cuda.exe cuda-release/voicebox-server-cuda-${PLATFORM}.exe
-          echo "Built CUDA server: voicebox-server-cuda-${PLATFORM}.exe (~3GB)"
+          echo "Built voicebox-server-${PLATFORM}.exe"

      - name: Setup Bun
        uses: oven-sh/setup-bun@v2
@@ -128,7 +106,7 @@ jobs:
      - name: Rust cache
        uses: swatinem/rust-cache@v2
        with:
-          workspaces: './tauri/src-tauri -> target'
+          workspaces: "./tauri/src-tauri -> target"

      - name: Install dependencies
        run: bun install
@@ -164,7 +142,7 @@ jobs:
        with:
          projectPath: tauri
          tagName: v__VERSION__
-          releaseName: 'voicebox v__VERSION__'
+          releaseName: "voicebox v__VERSION__"
          releaseBody: |
            ## What's Changed
            See the assets below to download and install this version.
@@ -172,41 +150,11 @@ jobs:
            ### Installation
            - **macOS (Apple Silicon)**: Download the `aarch64.dmg` file - uses MLX for fast native inference
            - **macOS (Intel)**: Download the `x64.dmg` file - uses PyTorch
-            - **Windows**: Download the `.msi` installer - includes CPU-only inference (~500MB)
+            - **Windows**: Download the `.msi` installer
            - **Linux**: Download the `.AppImage` or `.deb` package

-            ### NVIDIA GPU Acceleration (Windows)
-            Windows users with NVIDIA GPUs can enable CUDA for 4-5x faster inference:
-            1. Install the app normally (CPU version included in installer)
-            2. The app will detect your GPU and offer to download CUDA support automatically
-            3. Or manually download: [voicebox-server-cuda-x86_64-pc-windows-msvc.exe](https://downloads.voicebox.sh/cuda/__VERSION__/voicebox-server-cuda-x86_64-pc-windows-msvc.exe) (~2.4GB)
-
            The app includes automatic updates - future updates will be installed automatically.
          releaseDraft: true
          prerelease: false
          args: ${{ matrix.args }}
          includeUpdaterJson: true
-
-      - name: Upload CUDA server to Cloudflare R2 (Windows only)
-        if: matrix.platform == 'windows-latest'
-        env:
-          AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
-          AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
-          R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
-        run: |
-          # Install AWS CLI if not available
-          pip install awscli
-
-          # Get version from tag
-          VERSION=${GITHUB_REF#refs/tags/}
-
-          # Get platform tuple
-          PLATFORM=$(rustc --print host-tuple)
-
-          # Upload to R2
-          aws s3 cp backend/cuda-release/voicebox-server-cuda-${PLATFORM}.exe \
-            s3://voicebox/cuda/${VERSION}/voicebox-server-cuda-${PLATFORM}.exe \
-            --endpoint-url $R2_ENDPOINT \
-            --acl public-read
-
-          echo "CUDA binary uploaded to: https://downloads.voicebox.sh/cuda/${VERSION}/voicebox-server-cuda-${PLATFORM}.exe"
@@ -53,6 +53,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0

 ## [Unreleased]

+### Fixed
+- Audio export failing when Tauri save dialog returns object instead of string path
+
 ### Added
 - **Makefile** - Comprehensive development workflow automation with commands for setup, development, building, testing, and code quality checks
  - Includes Python version detection and compatibility warnings
@@ -59,7 +59,7 @@

 ## What is Voicebox?

-Voicebox is a **local-first voice cloning studio** with DAW-like features for professional voice synthesis. Think of it as the **Ollama for voice** — download models, clone voices, and generate speech entirely on your machine.
+Voicebox is a **local-first voice cloning studio** with DAW-like features for professional voice synthesis. Think of it as a **local, free and open-source alternative to ElevenLabs** — download models, clone voices, and generate speech entirely on your machine.

 Unlike cloud services that lock your voice data behind subscriptions, Voicebox gives you:

@@ -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.

@@ -233,7 +233,7 @@ See [CONTRIBUTING.md](CONTRIBUTING.md) for detailed setup and contribution guide

 ```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
@@ -247,7 +247,7 @@ make dev

 ```bash
 # Clone the repo
-git clone https://github.com/voicebox-sh/voicebox.git
+git clone https://github.com/jamiepine/voicebox.git
 cd voicebox

 # Install dependencies
@@ -260,7 +260,7 @@ cd backend && pip install -r requirements.txt && cd ..
 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
@@ -1,6 +1,6 @@
 {
  "name": "@voicebox/app",
-  "version": "0.1.12",
+  "version": "0.1.13",
  "private": true,
  "type": "module",
  "scripts": {
@@ -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>
@@ -1,6 +1,6 @@
 import { useMatchRoute } from '@tanstack/react-router';
 import { AnimatePresence, motion } from 'framer-motion';
-import { Loader2, MessageSquare, Sparkles } from 'lucide-react';
+import { Loader2, SlidersHorizontal, Sparkles } from 'lucide-react';
 import { useEffect, useRef, useState } from 'react';
 import { Button } from '@/components/ui/button';
 import { Form, FormControl, FormField, FormItem, FormMessage } from '@/components/ui/form';
@@ -187,7 +187,7 @@ export function FloatingGenerateBox({
      }}
    >
      <motion.div
-        className="bg-background/30 backdrop-blur-2xl border border-accent/20 rounded-[2rem] shadow-2xl hover:bg-background/40 hover:border-accent/20 transition-all duration-300 overflow-hidden p-3"
+        className="bg-background/30 backdrop-blur-2xl border border-accent/20 rounded-[2rem] shadow-2xl hover:bg-background/40 hover:border-accent/20 transition-all duration-300 p-3"
        transition={{ duration: 0.6, ease: 'easeInOut' }}
      >
        <Form {...form}>
@@ -274,7 +274,7 @@ export function FloatingGenerateBox({
                                  field.ref(node);
                                }
                              }}
-                              placeholder="Add delivery instructions..."
+                              placeholder="e.g. very happy and excited"
                              className="resize-none bg-transparent border-none focus-visible:ring-0 focus-visible:ring-offset-0 focus:outline-none focus:ring-0 outline-none ring-0 rounded-2xl text-sm placeholder:text-muted-foreground/60 w-full"
                              style={{
                                minHeight: isExpanded ? '100px' : '32px',
@@ -294,18 +294,27 @@ export function FloatingGenerateBox({
              </motion.div>

              <div className="relative shrink-0">
-                <Button
-                  type="submit"
-                  disabled={isPending || !selectedProfileId}
-                  className="h-10 w-10 rounded-full bg-accent hover:bg-accent/90 hover:scale-105 text-accent-foreground shadow-lg hover:shadow-accent/50 transition-all duration-200"
-                  size="icon"
-                >
-                  {isPending ? (
-                    <Loader2 className="h-4 w-4 animate-spin" />
-                  ) : (
-                    <Sparkles className="h-4 w-4" />
-                  )}
-                </Button>
+                <div className="group relative">
+                  <Button
+                    type="submit"
+                    disabled={isPending || !selectedProfileId}
+                    className="h-10 w-10 rounded-full bg-accent hover:bg-accent/90 hover:scale-105 text-accent-foreground shadow-lg hover:shadow-accent/50 transition-all duration-200"
+                    size="icon"
+                  >
+                    {isPending ? (
+                      <Loader2 className="h-4 w-4 animate-spin" />
+                    ) : (
+                      <Sparkles className="h-4 w-4" />
+                    )}
+                  </Button>
+                  <span className="pointer-events-none absolute bottom-full left-1/2 -translate-x-1/2 mb-2 whitespace-nowrap rounded-md bg-popover px-3 py-1.5 text-xs text-popover-foreground border border-border opacity-0 transition-opacity group-hover:opacity-100 z-[9999]">
+                    {isPending
+                      ? 'Generating...'
+                      : !selectedProfileId
+                        ? 'Select a voice profile first'
+                        : 'Generate speech'}
+                  </span>
+                </div>
                <AnimatePresence>
                  {isExpanded && (
                    <motion.div
@@ -315,20 +324,25 @@ export function FloatingGenerateBox({
                      transition={{ duration: 0.2 }}
                      className="absolute top-0 right-[calc(100%+0.5rem)]"
                    >
-                      <Button
-                        type="button"
-                        variant="ghost"
-                        size="icon"
-                        onClick={() => setIsInstructMode(!isInstructMode)}
-                        className={cn(
-                          'h-10 w-10 rounded-full transition-all duration-200',
-                          isInstructMode
-                            ? 'bg-accent text-accent-foreground border border-accent hover:bg-accent/90'
-                            : 'bg-card border border-border hover:bg-background/50',
-                        )}
-                      >
-                        <MessageSquare className="h-4 w-4" />
-                      </Button>
+                      <div className="group relative">
+                        <Button
+                          type="button"
+                          variant="ghost"
+                          size="icon"
+                          onClick={() => setIsInstructMode(!isInstructMode)}
+                          className={cn(
+                            'h-10 w-10 rounded-full transition-all duration-200',
+                            isInstructMode
+                              ? 'bg-accent text-accent-foreground border border-accent hover:bg-accent/90'
+                              : 'bg-card border border-border hover:bg-background/50',
+                          )}
+                        >
+                          <SlidersHorizontal className="h-4 w-4" />
+                        </Button>
+                        <span className="pointer-events-none absolute bottom-full left-1/2 -translate-x-1/2 mb-2 whitespace-nowrap rounded-md bg-popover px-3 py-1.5 text-xs text-popover-foreground border border-border opacity-0 transition-opacity group-hover:opacity-100 z-[9999]">
+                          Fine tune instructions
+                        </span>
+                      </div>
                    </motion.div>
                  )}
                </AnimatePresence>
@@ -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;

@@ -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,
          });

@@ -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>
@@ -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);
    }

@@ -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;
-  });
+  }
 }

 /**
@@ -1,3 +1,3 @@
 # Backend package

-__version__ = "0.1.12"
+__version__ = "0.1.13"
@@ -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)
@@ -358,9 +382,22 @@ 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:
@@ -5,7 +5,6 @@ PyInstaller build script for creating standalone Python server binary.
 import PyInstaller.__main__
 import os
 import platform
-import sys
 from pathlib import Path


@@ -14,27 +13,15 @@ def is_apple_silicon():
    return platform.system() == "Darwin" and platform.machine() == "arm64"


-def build_server(variant="cpu"):
-    """Build Python server as standalone binary.
-
-    Args:
-        variant: 'cpu' for CPU-only build (~500MB) or 'cuda' for CUDA build (~3GB)
-    """
+def build_server():
+    """Build Python server as standalone binary."""
    backend_dir = Path(__file__).parent

-    if variant not in ['cpu', 'cuda']:
-        raise ValueError(f"Invalid variant: {variant}. Must be 'cpu' or 'cuda'")
-
-    # Set binary name based on variant
-    binary_name = f'voicebox-server-{variant}' if variant == 'cuda' else 'voicebox-server'
-
-    print(f"Building {variant.upper()} variant: {binary_name}")
-
    # PyInstaller arguments
    args = [
        'server.py',  # Use server.py as entry point instead of main.py
        '--onefile',
-        '--name', binary_name,
+        '--name', 'voicebox-server',
    ]

    # Add local qwen_tts path if specified (for editable installs)
@@ -96,9 +83,13 @@ def build_server(variant="cpu"):
            '--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:
        print("Building for non-Apple Silicon platform - PyTorch only")
@@ -113,14 +104,9 @@ def build_server(variant="cpu"):
    
    # Run PyInstaller
    PyInstaller.__main__.run(args)
-
-    print(f"\n{'='*60}")
-    print(f"Build complete: {variant.upper()} variant")
-    print(f"Binary: {backend_dir / 'dist' / binary_name}")
-    print(f"{'='*60}\n")
+    
+    print(f"Binary built in {backend_dir / 'dist' / 'voicebox-server'}")


 if __name__ == '__main__':
-    # Accept variant as command line argument
-    variant = sys.argv[1] if len(sys.argv) > 1 else 'cpu'
-    build_server(variant)
+    build_server()
@@ -1,30 +0,0 @@
-@echo off
-REM Build both CPU and CUDA server binaries for Windows
-
-echo ============================================================
-echo Building BOTH server binaries (CPU + CUDA)
-echo This will take a while...
-echo ============================================================
-
-call build_cpu.bat
-if errorlevel 1 (
-    echo CPU build failed!
-    exit /b 1
-)
-
-echo.
-echo.
-
-call build_cuda.bat
-if errorlevel 1 (
-    echo CUDA build failed!
-    exit /b 1
-)
-
-echo.
-echo ============================================================
-echo Both binaries built successfully!
-echo ============================================================
-echo CPU binary:  dist\voicebox-server.exe (~500MB)
-echo CUDA binary: dist\voicebox-server-cuda.exe (~3GB)
-echo ============================================================
@@ -1,28 +0,0 @@
-@echo off
-REM Build CPU-only server binary for Windows
-REM This creates a ~500MB binary without CUDA support
-
-echo ============================================================
-echo Building CPU-only server binary
-echo ============================================================
-
-echo.
-echo Step 1: Installing CPU-only PyTorch...
-pip uninstall -y torch torchvision torchaudio
-pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-
-echo.
-echo Step 2: Building binary with PyInstaller...
-python build_binary.py cpu
-
-echo.
-echo Step 3: Restoring CUDA PyTorch for development...
-pip uninstall -y torch torchvision torchaudio
-pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
-
-echo.
-echo ============================================================
-echo CPU binary built successfully!
-echo Location: dist\voicebox-server.exe
-echo Size: ~500MB
-echo ============================================================
@@ -1,30 +0,0 @@
-#!/bin/bash
-# Build CPU-only server binary
-# This creates a ~500MB binary without CUDA support
-
-set -e
-
-echo "============================================================"
-echo "Building CPU-only server binary"
-echo "============================================================"
-
-echo ""
-echo "Step 1: Installing CPU-only PyTorch..."
-pip uninstall -y torch torchvision torchaudio || true
-pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-
-echo ""
-echo "Step 2: Building binary with PyInstaller..."
-python build_binary.py cpu
-
-echo ""
-echo "Step 3: Restoring CUDA PyTorch for development..."
-pip uninstall -y torch torchvision torchaudio || true
-pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
-
-echo ""
-echo "============================================================"
-echo "CPU binary built successfully!"
-echo "Location: dist/voicebox-server"
-echo "Size: ~500MB"
-echo "============================================================"
@@ -1,22 +0,0 @@
-@echo off
-REM Build CUDA server binary for Windows
-REM This creates a ~3GB binary with CUDA support
-
-echo ============================================================
-echo Building CUDA server binary
-echo ============================================================
-
-echo.
-echo Step 1: Ensuring CUDA PyTorch is installed...
-pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 --upgrade
-
-echo.
-echo Step 2: Building binary with PyInstaller...
-python build_binary.py cuda
-
-echo.
-echo ============================================================
-echo CUDA binary built successfully!
-echo Location: dist\voicebox-server-cuda.exe
-echo Size: ~3GB
-echo ============================================================
@@ -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")

@@ -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
@@ -77,10 +95,39 @@ 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 +136,10 @@ 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:
@@ -252,12 +303,17 @@ 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 +324,8 @@ 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 +446,7 @@ 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:
@@ -542,47 +600,50 @@ 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
+
+        # Resolve model size and load the correct model FIRST.
+        # This must happen before create_voice_prompt_for_profile because that
+        # function calls load_model_async(None), which falls back to self.model_size.
+        # If the model is already loaded with the right size at that point, it
+        # returns immediately and the voice prompt is created by the correct model.
        tts_model = tts.get_tts_model()
-        # Load the requested model size if different from current (async to not block)
        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
-                model_name = f"qwen-tts-{model_size}"
+        if not tts_model._is_model_cached(model_size):
+            # Model is not fully cached — kick off a background download and tell
+            # the client to retry once it's ready.
+            model_name = f"qwen-tts-{model_size}"

-                async def download_model_background():
-                    try:
-                        await tts_model.load_model_async(model_size)
-                    except Exception as e:
-                        task_manager.error_download(model_name, str(e))
+            async def download_model_background():
+                try:
+                    await tts_model.load_model_async(model_size)
+                except Exception as e:
+                    task_manager.error_download(model_name, str(e))

-                task_manager.start_download(model_name)
-                asyncio.create_task(download_model_background())
+            task_manager.start_download(model_name)
+            asyncio.create_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
-                    }
-                )
+            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,
+                },
+            )

+        # Load (or switch to) the requested model before building the voice prompt
        await tts_model.load_model_async(model_size)
+
+        # Create voice prompt from profile (model is already loaded with correct size)
+        voice_prompt = await profiles.create_voice_prompt_for_profile(
+            data.profile_id,
+            db,
+        )
+
        audio, sample_rate = await tts_model.generate(
            data.text,
            voice_prompt,
@@ -625,6 +686,59 @@ 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.
+    """
+    profile = await profiles.get_profile(data.profile_id, db)
+    if not profile:
+        raise HTTPException(status_code=404, detail="Profile not found")
+
+    tts_model = tts.get_tts_model()
+    model_size = data.model_size or "1.7B"
+
+    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.",
+        )
+
+    # Load the correct model before building the voice prompt (fixes issue #96)
+    await tts_model.load_model_async(model_size)
+
+    voice_prompt = await profiles.create_voice_prompt_for_profile(data.profile_id, db)
+
+    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 +867,7 @@ 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 +900,7 @@ 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)
        }
    )

@@ -1054,7 +1168,7 @@ 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:
@@ -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"
@@ -18,6 +18,7 @@ qwen-tts>=0.0.5
 librosa>=0.10.0
 soundfile>=0.12.0
 numpy>=1.24.0
+numba>=0.60.0,<0.61.0

 # Utilities
 python-multipart>=0.0.6
@@ -1,137 +0,0 @@
-#!/usr/bin/env python3
-"""
-Test CUDA binary compression to verify it fits under GitHub's 2GB release asset limit.
-
-Usage:
-    python test_cuda_compression.py [path/to/voicebox-server-cuda.exe]
-
-If no path provided, looks for the binary in ./dist/
-"""
-
-import os
-import sys
-import subprocess
-from pathlib import Path
-
-
-def format_size(bytes_size):
-    """Format bytes into human-readable size."""
-    for unit in ['B', 'KB', 'MB', 'GB']:
-        if bytes_size < 1024.0:
-            return f"{bytes_size:.2f} {unit}"
-        bytes_size /= 1024.0
-    return f"{bytes_size:.2f} TB"
-
-
-def get_file_size(filepath):
-    """Get file size in bytes."""
-    return os.path.getsize(filepath)
-
-
-def compress_with_7z(input_file, output_file):
-    """Compress file using 7z with maximum compression."""
-    print(f"\nCompressing with 7z (maximum compression)...")
-    print(f"This may take several minutes for a ~2.5GB file...\n")
-
-    cmd = [
-        '7z', 'a',
-        '-t7z',           # 7z format
-        '-m0=lzma2',      # LZMA2 compression
-        '-mx=9',          # Maximum compression
-        '-mfb=64',        # Fast bytes
-        '-md=32m',        # Dictionary size
-        '-ms=on',         # Solid archive
-        output_file,
-        input_file
-    ]
-
-    try:
-        subprocess.run(cmd, check=True, capture_output=True, text=True)
-        return True
-    except subprocess.CalledProcessError as e:
-        print(f"Error during compression: {e}")
-        print(f"stderr: {e.stderr}")
-        return False
-    except FileNotFoundError:
-        print("ERROR: 7z not found. Please install 7-Zip:")
-        print("  Windows: https://www.7-zip.org/download.html")
-        print("  macOS: brew install p7zip")
-        print("  Linux: apt-get install p7zip-full")
-        return False
-
-
-def main():
-    # Find CUDA binary
-    if len(sys.argv) > 1:
-        cuda_binary = Path(sys.argv[1])
-    else:
-        # Look in dist directory
-        dist_dir = Path(__file__).parent / 'dist'
-        candidates = list(dist_dir.glob('voicebox-server-cuda*.exe'))
-
-        if not candidates:
-            print("ERROR: CUDA binary not found in ./dist/")
-            print("Please provide the path as an argument:")
-            print("  python test_cuda_compression.py path/to/voicebox-server-cuda.exe")
-            sys.exit(1)
-
-        cuda_binary = candidates[0]
-
-    if not cuda_binary.exists():
-        print(f"ERROR: File not found: {cuda_binary}")
-        sys.exit(1)
-
-    print("=" * 70)
-    print("CUDA Binary Compression Test")
-    print("=" * 70)
-
-    # Get original size
-    original_size = get_file_size(cuda_binary)
-    print(f"\nOriginal file: {cuda_binary.name}")
-    print(f"Original size: {format_size(original_size)} ({original_size:,} bytes)")
-
-    # Check if already over 2GB
-    github_limit = 2 * 1024 * 1024 * 1024  # 2GB in bytes
-    print(f"GitHub limit:  {format_size(github_limit)} ({github_limit:,} bytes)")
-
-    if original_size > github_limit:
-        print(f"\n[WARNING] Original file exceeds GitHub limit by {format_size(original_size - github_limit)}")
-    else:
-        print(f"\n[OK] Original file is under GitHub limit")
-
-    # Compress
-    output_file = cuda_binary.parent / f"{cuda_binary.stem}.7z"
-    if output_file.exists():
-        print(f"\nRemoving existing compressed file: {output_file.name}")
-        output_file.unlink()
-
-    success = compress_with_7z(cuda_binary, output_file)
-
-    if not success:
-        sys.exit(1)
-
-    # Check compressed size
-    compressed_size = get_file_size(output_file)
-    compression_ratio = (1 - compressed_size / original_size) * 100
-
-    print("\n" + "=" * 70)
-    print("Compression Results")
-    print("=" * 70)
-    print(f"\nCompressed file: {output_file.name}")
-    print(f"Compressed size: {format_size(compressed_size)} ({compressed_size:,} bytes)")
-    print(f"Compression ratio: {compression_ratio:.1f}%")
-    print(f"Space saved: {format_size(original_size - compressed_size)}")
-
-    if compressed_size <= github_limit:
-        print(f"\n[SUCCESS] Compressed file fits under GitHub's 2GB limit!")
-        print(f"   Margin: {format_size(github_limit - compressed_size)} remaining")
-    else:
-        print(f"\n[FAILED] Compressed file still exceeds GitHub limit")
-        print(f"   Over by: {format_size(compressed_size - github_limit)}")
-        print(f"\n   Alternative: Host on external storage (S3, Azure Blob, etc.)")
-
-    print("\n" + "=" * 70)
-
-
-if __name__ == '__main__':
-    main()
@@ -1,86 +0,0 @@
-#!/bin/bash
-# Test R2 upload locally before running in CI
-
-set -e
-
-echo "============================================================"
-echo "Cloudflare R2 Upload Test"
-echo "============================================================"
-
-# Check for required environment variables
-if [ -z "$AWS_ACCESS_KEY_ID" ] || [ -z "$AWS_SECRET_ACCESS_KEY" ] || [ -z "$R2_ENDPOINT" ]; then
-    echo "ERROR: Missing required environment variables"
-    echo ""
-    echo "Please set:"
-    echo "  export AWS_ACCESS_KEY_ID='your-r2-access-key-id'"
-    echo "  export AWS_SECRET_ACCESS_KEY='your-r2-secret-access-key'"
-    echo "  export R2_ENDPOINT='https://your-account-id.r2.cloudflarestorage.com'"
-    echo ""
-    exit 1
-fi
-
-# Check for AWS CLI
-if ! command -v aws &> /dev/null; then
-    echo "Installing AWS CLI..."
-    pip install awscli
-fi
-
-# Find CUDA binary
-CUDA_BINARY=$(ls dist/voicebox-server-cuda*.exe 2>/dev/null | head -1)
-
-if [ -z "$CUDA_BINARY" ]; then
-    echo "ERROR: CUDA binary not found in dist/"
-    echo "Run: bash build_cuda.bat"
-    exit 1
-fi
-
-echo ""
-echo "Found CUDA binary: $CUDA_BINARY"
-echo "Size: $(du -h "$CUDA_BINARY" | cut -f1)"
-echo ""
-
-# Test version
-VERSION="v0.1.12-test"
-PLATFORM="x86_64-pc-windows-msvc"
-FILENAME="voicebox-server-cuda-${PLATFORM}.exe"
-
-echo "Test upload configuration:"
-echo "  Version: $VERSION"
-echo "  Platform: $PLATFORM"
-echo "  Endpoint: $R2_ENDPOINT"
-echo "  Bucket: voicebox"
-echo "  Path: cuda/$VERSION/$FILENAME"
-echo ""
-
-read -p "Proceed with upload? (y/n) " -n 1 -r
-echo
-if [[ ! $REPLY =~ ^[Yy]$ ]]; then
-    echo "Aborted."
-    exit 1
-fi
-
-echo ""
-echo "Uploading to R2..."
-
-aws s3 cp "$CUDA_BINARY" \
-    "s3://voicebox/cuda/${VERSION}/${FILENAME}" \
-    --endpoint-url "$R2_ENDPOINT" \
-    --acl public-read
-
-if [ $? -eq 0 ]; then
-    echo ""
-    echo "============================================================"
-    echo "Upload successful!"
-    echo "============================================================"
-    echo ""
-    echo "Download URL:"
-    echo "https://downloads.voicebox.sh/cuda/${VERSION}/${FILENAME}"
-    echo ""
-    echo "Test with:"
-    echo "curl -I https://downloads.voicebox.sh/cuda/${VERSION}/${FILENAME}"
-    echo ""
-else
-    echo ""
-    echo "Upload failed!"
-    exit 1
-fi
@@ -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()
@@ -1,48 +0,0 @@
-# -*- mode: python ; coding: utf-8 -*-
-from PyInstaller.utils.hooks import collect_data_files
-from PyInstaller.utils.hooks import collect_submodules
-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']
-datas += collect_data_files('qwen_tts')
-datas += copy_metadata('qwen-tts')
-hiddenimports += collect_submodules('qwen_tts')
-hiddenimports += collect_submodules('jaraco')
-
-
-a = Analysis(
-    ['server.py'],
-    pathex=[],
-    binaries=[],
-    datas=datas,
-    hiddenimports=hiddenimports,
-    hookspath=[],
-    hooksconfig={},
-    runtime_hooks=[],
-    excludes=[],
-    noarchive=False,
-    optimize=0,
-)
-pyz = PYZ(a.pure)
-
-exe = EXE(
-    pyz,
-    a.scripts,
-    a.binaries,
-    a.datas,
-    [],
-    name='voicebox-server-cuda',
-    debug=False,
-    bootloader_ignore_signals=False,
-    strip=False,
-    upx=True,
-    upx_exclude=[],
-    runtime_tmpdir=None,
-    console=True,
-    disable_windowed_traceback=False,
-    argv_emulation=False,
-    target_arch=None,
-    codesign_identity=None,
-    entitlements_file=None,
-)
@@ -4,17 +4,26 @@ from PyInstaller.utils.hooks import collect_submodules
 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']
+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')
+# 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')
+hiddenimports += collect_submodules('mlx')
+hiddenimports += collect_submodules('mlx_audio')


 a = Analysis(
    ['server.py'],
    pathex=[],
-    binaries=[],
+    binaries=_mlx_bins + _mlxa_bins,
    datas=datas,
    hiddenimports=hiddenimports,
    hookspath=[],
@@ -1,620 +0,0 @@
-# CUDA Distribution Problem - Complete Analysis
-
-## Table of Contents
-1. [Problem Overview](#problem-overview)
-2. [Root Cause](#root-cause)
-3. [Attempted Solutions](#attempted-solutions)
-4. [Current Status](#current-status)
-5. [Available Options](#available-options)
-6. [Technical Details](#technical-details)
-7. [Cost Analysis](#cost-analysis)
-8. [Recommendations](#recommendations)
-
---
-
-## Problem Overview
-
-### Timeline of Issues
-
-**Original Problem (v0.1.0 - v0.1.11)**
- Single server binary with CUDA support
- Size: ~2.9GB
- Issue: MSI installer build fails in GitHub Actions CI
- Error: WiX Toolset cannot handle 3GB files efficiently
-
-**First Solution: Dual Binary System (v0.1.12)**
- Split into CPU (295MB) and CUDA (2.37GB) binaries
- CPU ships with installer
- CUDA as optional download
- Issue: GitHub Release assets have 2GB limit
-
-**Current Problem (Discovered during implementation)**
- GitHub Release Asset Limit: **2GB hard maximum**
- CUDA binary: **2.37GB** (370MB over limit)
- Cannot upload to GitHub Releases
-
---
-
-## Root Cause
-
-### Why Is The CUDA Binary So Large?
-
-The size difference between CPU and CUDA builds:
-
-| Component | CPU Build | CUDA Build | Difference |
-|-----------|-----------|------------|------------|
-| PyTorch Core | ~150MB | ~150MB | - |
-| CPU Libraries (MKL/OpenBLAS) | ~100MB | - | -100MB |
-| CUDA Runtime | - | ~500MB | +500MB |
-| cuBLAS | - | ~350MB | +350MB |
-| cuDNN | - | ~1.2GB | +1.2GB |
-| NVRTC (CUDA Compiler) | - | ~90MB | +90MB |
-| Other CUDA libs | - | ~100MB | +100MB |
-| **Total** | **~295MB** | **~2.37GB** | **+2.07GB** |
-
-### CUDA Dependencies Breakdown
-
-```
-torch/lib/ (CUDA build):
-├── cudart64_12.dll          (~0.5 MB)   - CUDA Runtime
-├── cublas64_12.dll          (~100 MB)   - Basic Linear Algebra
-├── cublasLt64_12.dll        (~200 MB)   - Linear Algebra (optimized)
-├── cudnn64_9.dll            (~800 MB)   - Deep Neural Networks
-├── cudnn_*_infer64_9.dll    (~400 MB)   - DNN Inference ops
-├── nvrtc64_*.dll            (~50 MB)    - Runtime Compiler
-├── nvrtc-builtins64_*.dll   (~40 MB)    - Compiler builtins
-├── torch_cuda.dll           (~200 MB)   - PyTorch CUDA bridge
-└── c10_cuda.dll             (~20 MB)    - Core CUDA utilities
-```
-
-**Why These Are Required:**
- cuDNN is essential for neural network operations
- cuBLAS handles all matrix operations (core of ML)
- Cannot split or remove without breaking functionality
-
---
-
-## Attempted Solutions
-
-### Solution 1: Dual Binary System ✅ (Partially Successful)
-
-**Goal**: Split CPU and CUDA into separate downloads
-
-**Implementation**:
-```bash
-# Build CPU-only (295MB)
-pip install torch --index-url https://download.pytorch.org/whl/cpu
-python build_binary.py cpu
-
-# Build CUDA (2.37GB)
-pip install torch --index-url https://download.pytorch.org/whl/cu121
-python build_binary.py cuda
-```
-
-**Results**:
- ✅ CPU binary: 295MB (fits in installer)
- ✅ CI builds successfully
- ✅ Installer size reduced from 3GB to ~500MB
- ❌ CUDA binary still too large for GitHub
-
-**See**: `docs/dual-server-binaries.md`
-
-### Solution 2: Compression Testing ❌ (Failed)
-
-**Goal**: Compress CUDA binary to fit under 2GB
-
-**Method**: 7z with maximum compression settings
-```bash
-7z a -t7z -m0=lzma2 -mx=9 -mfb=64 -md=32m -ms=on \
-  voicebox-server-cuda.7z voicebox-server-cuda.exe
-```
-
-**Results**:
-```
-Original:     2.37 GB (2,545,086,396 bytes)
-Compressed:   2.35 GB (2,519,381,264 bytes)
-Compression:  1.0% (only 24.5MB saved)
-GitHub Limit: 2.00 GB (2,147,483,648 bytes)
-Over by:      354.67 MB
-
-Status: FAILED - Still exceeds limit by 354MB
-```
-
-**Why Compression Failed**:
- CUDA binaries are already optimized machine code
- No redundant data to compress
- Neural network kernels are highly compact
- Libraries are already stripped of debug symbols
-
-**Conclusion**: Compression is not viable
-
---
-
-## Current Status
-
-### What Works
- ✅ CPU binary builds successfully (295MB)
- ✅ CUDA binary builds successfully (2.37GB)
- ✅ Build scripts for both variants
- ✅ CI workflow updated for dual binaries
- ✅ Installer can be created with CPU binary
-
-### What Doesn't Work
- ❌ Cannot upload CUDA binary to GitHub Releases (exceeds 2GB limit)
- ❌ Compression doesn't reduce size enough
- ❌ No automated distribution path for CUDA binary
-
-### Branch Status
- Branch: `feat/dual-server-binaries`
- Commits: Implementation complete
- Testing: Local builds successful
- Blocker: CUDA distribution path
-
---
-
-## Available Options
-
-### Option 1: AWS S3 Hosting (Recommended)
-
-**Description**: Host CUDA binary in Amazon S3 bucket
-
-**Pros**:
- ✅ No file size limits (can handle multi-GB files)
- ✅ Fast global CDN (CloudFront)
- ✅ Reliable (99.99% uptime)
- ✅ Pay only for usage
- ✅ Easy CI integration
- ✅ Version control (keep multiple releases)
-
-**Cons**:
- ❌ Requires AWS account
- ❌ Monthly costs (~$1-5/month)
- ❌ Additional infrastructure to manage
-
-**Cost Estimate**:
-```
-Storage: 2.37 GB × $0.023/GB = $0.05/month
-Transfer: 100 downloads × 2.37GB × $0.09/GB = $21.33/month
-Total: ~$21-25/month for 100 downloads
-      ~$2-5/month for 10-20 downloads
-```
-
-**Implementation**:
-```yaml
-# .github/workflows/release.yml
- name: Upload CUDA to S3
-  env:
-    AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
-    AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
-  run: |
-    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
-      s3://voicebox-releases/cuda/${{ github.ref_name }}/ \
-      --acl public-read
-
-    # Generate download URL
-    echo "CUDA_URL=https://voicebox-releases.s3.amazonaws.com/cuda/${{ github.ref_name }}/voicebox-server-cuda-x86_64-pc-windows-msvc.exe" >> release_notes.txt
-```
-
-**User Experience**:
-1. Install app normally (500MB installer)
-2. App detects NVIDIA GPU
-3. Shows: "Download CUDA support? (2.4GB)"
-4. Downloads from S3: `https://voicebox-releases.s3.amazonaws.com/cuda/v0.1.12/voicebox-server-cuda.exe`
-5. Saves to `%APPDATA%/voicebox/binaries/`
-6. App restarts with CUDA server
-
---
-
-### Option 2: Azure Blob Storage
-
-**Description**: Microsoft Azure alternative to S3
-
-**Pros**:
- ✅ Similar to S3 (no size limits, CDN, reliable)
- ✅ Good if already using Azure
- ✅ Competitive pricing
- ✅ Global CDN with Azure CDN
-
-**Cons**:
- ❌ Requires Azure account
- ❌ Similar monthly costs
- ❌ Less common in open source projects
-
-**Cost Estimate**:
-```
-Storage: $0.018/GB = $0.04/month
-Transfer: ~$20-25/month for 100 downloads
-```
-
-**Implementation**:
-```yaml
- name: Upload to Azure Blob
-  env:
-    AZURE_STORAGE_CONNECTION_STRING: ${{ secrets.AZURE_STORAGE }}
-  run: |
-    az storage blob upload \
-      --account-name voiceboxreleases \
-      --container-name cuda-binaries \
-      --name v${{ github.ref_name }}/voicebox-server-cuda.exe \
-      --file backend/cuda-release/voicebox-server-cuda-*.exe \
-      --tier Hot
-```
-
---
-
-### Option 3: Cloudflare R2
-
-**Description**: Cloudflare's S3-compatible object storage
-
-**Pros**:
- ✅ S3-compatible API
- ✅ **FREE egress (no bandwidth charges!)**
- ✅ Cheaper than S3/Azure
- ✅ Cloudflare CDN included
- ✅ Good for open source projects
-
-**Cons**:
- ❌ Requires Cloudflare account
- ❌ Newer service (less mature than S3)
-
-**Cost Estimate**:
-```
-Storage: $0.015/GB = $0.04/month
-Egress: $0.00 (FREE!)
-Class A ops: Negligible
-Total: ~$0.04/month (essentially free!)
-```
-
-**Why This Is Attractive**:
- Zero bandwidth costs (huge savings)
- Perfect for open source distribution
- S3-compatible (easy migration if needed)
-
-**Implementation**:
-Same as S3 (R2 is S3-compatible):
-```yaml
- name: Upload to R2
-  env:
-    AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
-    AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
-    AWS_ENDPOINT_URL: https://<account-id>.r2.cloudflarestorage.com
-  run: |
-    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
-      s3://voicebox-releases/cuda/${{ github.ref_name }}/ \
-      --endpoint-url=$AWS_ENDPOINT_URL
-```
-
---
-
-### Option 4: GitHub Packages (Container Registry)
-
-**Description**: Package CUDA binary as OCI/Docker artifact
-
-**Pros**:
- ✅ Stays in GitHub ecosystem
- ✅ No additional accounts needed
- ✅ Free for public repos
-
-**Cons**:
- ❌ Complex for desktop app distribution
- ❌ Users need to extract from container
- ❌ Awkward UX (not designed for binary distribution)
- ❌ Requires Docker understanding
-
-**Not Recommended**: Containers aren't designed for desktop app binaries
-
---
-
-### Option 5: Self-Hosted Server
-
-**Description**: Host on your own VPS/server
-
-**Pros**:
- ✅ Full control
- ✅ No cloud provider dependency
- ✅ Predictable costs
-
-**Cons**:
- ❌ Requires server maintenance
- ❌ Bandwidth costs can be high
- ❌ Uptime responsibility
- ❌ Scaling challenges
-
-**Cost Estimate**:
-```
-VPS: $5-20/month (DigitalOcean, Linode)
-Bandwidth: $0.01-0.02/GB
-Total: $10-50/month depending on traffic
-```
-
---
-
-### Option 6: Manual Distribution
-
-**Description**: Don't automate - provide manual download instructions
-
-**Pros**:
- ✅ Zero cost
- ✅ Zero infrastructure
- ✅ Simple
-
-**Cons**:
- ❌ Poor user experience
- ❌ Manual upload to file host each release
- ❌ Users must manually download and install
- ❌ No automatic updates for CUDA binary
- ❌ Increases support burden
-
-**Implementation**:
-```
-Release notes:
-"Windows users with NVIDIA GPUs can download CUDA support:
-1. Download voicebox-server-cuda.exe from [Google Drive/Mega/etc]
-2. Place in C:\Users\<YourName>\AppData\Roaming\voicebox\binaries\
-3. Restart the app"
-```
-
-**Not Recommended**: Creates friction, support issues
-
---
-
-### Option 7: Split CUDA Binary
-
-**Description**: Break CUDA binary into multiple <2GB chunks
-
-**Technical Approach**:
-```python
-# Split binary
-split -b 2000M voicebox-server-cuda.exe cuda_part_
-
-# Upload parts to GitHub (each <2GB)
-cuda_part_aa (2.0 GB)
-cuda_part_ab (0.37 GB)
-
-# App downloads and reassembles
-cat cuda_part_* > voicebox-server-cuda.exe
-```
-
-**Pros**:
- ✅ Stays on GitHub
- ✅ No external hosting
-
-**Cons**:
- ❌ Complex download logic (multiple files)
- ❌ Integrity checking required
- ❌ More points of failure
- ❌ Users must wait for multiple downloads
- ❌ Still hacky solution
-
-**Complexity**: Medium-High
-
---
-
-## Technical Details
-
-### Current Build Output
-
-```
-backend/dist/
-├── voicebox-server.exe          295 MB  (CPU-only)
-└── voicebox-server-cuda.exe    2.37 GB  (CUDA)
-
-# After compression test:
-backend/dist/
-└── voicebox-server-cuda.7z     2.35 GB  (not viable)
-```
-
-### CI Workflow Changes Required
-
-For external hosting (S3/R2/Azure):
-
-```yaml
-# Current workflow (fails)
- name: Upload CUDA server binary (Windows only)
-  if: matrix.platform == 'windows-latest'
-  uses: softprops/action-gh-release@v1
-  with:
-    files: backend/cuda-release/voicebox-server-cuda-*.exe  # ❌ Fails: >2GB
-    draft: true
-
-# New workflow (S3 example)
- name: Upload CUDA to S3 (Windows only)
-  if: matrix.platform == 'windows-latest'
-  env:
-    AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
-    AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
-  run: |
-    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
-      s3://voicebox-releases/cuda/${{ github.ref_name }}/ \
-      --acl public-read
-
-    # Generate release notes with download URL
-    cat >> release_notes.md <<EOF
-
-    ### GPU Acceleration (Windows)
-    Download CUDA support for NVIDIA GPUs:
-    [voicebox-server-cuda.exe](https://voicebox-releases.s3.amazonaws.com/cuda/${{ github.ref_name }}/voicebox-server-cuda-x86_64-pc-windows-msvc.exe)
-    Size: 2.37 GB
-    EOF
-```
-
-### App Changes Required
-
-**Frontend (Tauri)**: Download manager
-```typescript
-// src/lib/cuda-downloader.ts
-const CUDA_DOWNLOAD_URL =
-  "https://voicebox-releases.s3.amazonaws.com/cuda/v{VERSION}/voicebox-server-cuda.exe";
-
-async function downloadCudaBinary(version: string) {
-  const url = CUDA_DOWNLOAD_URL.replace("{VERSION}", version);
-  const savePath = path.join(app.getPath("userData"), "binaries", "voicebox-server-cuda.exe");
-
-  // Download with progress
-  await downloadFile(url, savePath, (progress) => {
-    // Update UI: "Downloading CUDA support: 45% (1.2GB / 2.4GB)"
-  });
-
-  // Verify checksum
-  const checksum = await calculateChecksum(savePath);
-  if (checksum !== EXPECTED_CHECKSUM) {
-    throw new Error("Download corrupted");
-  }
-}
-```
-
-**Backend**: Already supports both binaries (no changes needed)
-
---
-
-## Cost Analysis
-
-### Monthly Cost Comparison (100 downloads/month)
-
-| Option | Storage | Bandwidth | Total/Month | Notes |
-|--------|---------|-----------|-------------|-------|
-| **Cloudflare R2** | $0.04 | $0.00 | **$0.04** | Best for open source |
-| AWS S3 | $0.05 | $21.33 | $21.38 | Good reliability |
-| Azure Blob | $0.04 | $20.00 | $20.04 | Azure ecosystem |
-| Self-hosted VPS | $10.00 | $2.37 | $12.37 | Maintenance overhead |
-| Manual | $0.00 | $0.00 | $0.00 | Poor UX |
-
-### Annual Cost Comparison
-
-| Option | Year 1 | Year 2+ | Notes |
-|--------|--------|---------|-------|
-| **Cloudflare R2** | **$0.50** | **$0.50** | Essentially free |
-| AWS S3 | $256 | $256 | Predictable |
-| Self-hosted | $144 | $144 | Time cost |
-
-**Recommendation**: Cloudflare R2 (free egress = huge savings)
-
---
-
-## Recommendations
-
-### Recommended Solution: Cloudflare R2
-
-**Why**:
-1. **Cost**: Essentially free (~$0.04/month)
-2. **Bandwidth**: Zero egress charges (unlimited downloads)
-3. **CDN**: Cloudflare's global network included
-4. **Compatibility**: S3-compatible API (easy to use)
-5. **Perfect for open source**: No surprise bandwidth bills
-
-### Implementation Priority
-
-**Phase 1: Setup (1-2 hours)**
-1. Create Cloudflare R2 account
-2. Create bucket: `voicebox-releases`
-3. Generate API credentials
-4. Add to GitHub Secrets
-
-**Phase 2: CI Integration (1-2 hours)**
-1. Update `.github/workflows/release.yml`
-2. Add R2 upload step
-3. Generate release notes with download URL
-4. Test with draft release
-
-**Phase 3: App Integration (4-6 hours)**
-1. Add GPU detection on startup
-2. Implement download manager UI
-3. Add progress indicators
-4. Implement checksum verification
-5. Server restart logic
-
-**Phase 4: Documentation (1 hour)**
-1. Update README with GPU instructions
-2. Add troubleshooting guide
-3. Document manual download process
-
-**Total Time**: ~8-12 hours of development
-
-### Alternative: AWS S3 (If Already Using AWS)
-
-If you're already using AWS for other infrastructure, S3 is also a solid choice:
- More mature than R2
- Extensive documentation
- Familiar tooling
- ~$20/month for moderate usage
-
---
-
-## Open Questions
-
-1. **Expected Download Volume**: How many CUDA downloads per month?
-   - Affects cost calculations
-   - Determines if R2's free egress is significant
-
-2. **Update Strategy**: How to handle CUDA updates?
-   - Option A: Version in URL path (keep all versions)
-   - Option B: Overwrite latest (save space)
-
-3. **Fallback Strategy**: What if cloud provider is down?
-   - Mirror on multiple providers?
-   - Graceful degradation to CPU?
-
-4. **Telemetry**: Track CUDA download stats?
-   - Helps with cost forecasting
-   - User behavior insights
-
---
-
-## Next Steps
-
-1. **Research Phase** (You are here)
-   - Evaluate cloud providers
-   - Check terms of service
-   - Test account creation
-
-2. **Decision Phase**
-   - Choose provider (Cloudflare R2 recommended)
-   - Set up account
-   - Configure billing alerts
-
-3. **Implementation Phase**
-   - Update CI workflow
-   - Implement download manager
-   - Test end-to-end flow
-
-4. **Launch Phase**
-   - Deploy to production
-   - Monitor downloads
-   - Gather user feedback
-
---
-
-## References
-
- **GitHub Release Limits**: https://docs.github.com/en/repositories/releasing-projects-on-github/about-releases
- **Cloudflare R2 Pricing**: https://developers.cloudflare.com/r2/pricing/
- **AWS S3 Pricing**: https://aws.amazon.com/s3/pricing/
- **Compression Test Results**: `backend/test_cuda_compression.py`
- **Dual Binary Implementation**: `docs/dual-server-binaries.md`
-
---
-
-## Appendix: Alternative Approaches Considered
-
-### A. Dynamic CUDA Loading
-**Idea**: Load CUDA DLLs dynamically at runtime
-**Why Not**: PyTorch requires CUDA DLLs at import time, can't lazy-load
-
-### B. CUDA as Separate Package
-**Idea**: Python package with just CUDA libs
-**Why Not**: Still 2GB+, same problem
-
-### C. Model Quantization
-**Idea**: Use smaller quantized models
-**Why Not**: Doesn't reduce CUDA runtime size
-
-### D. Docker Distribution
-**Idea**: Distribute as Docker container
-**Why Not**: Poor fit for desktop app, requires Docker installed
-
---
-
-**Document Version**: 1.0
-**Last Updated**: 2026-01-31
-**Status**: Research Phase
-**Next Review**: After cloud provider decision
@@ -1,177 +0,0 @@
-# Dual Server Binary System
-
-## Overview
-
-Voicebox now uses a dual-binary approach to manage the size difference between CPU-only and CUDA-enabled builds:
-
- **CPU Binary** (~500MB): Ships with the installer by default
- **CUDA Binary** (~3GB): Downloaded on-demand for GPU users
-
-## Problem Solved
-
-Previously, bundling PyTorch with CUDA support created a 3GB server binary, which:
- Made the installer too large (failed CI builds with WiX)
- Forced all users to download CUDA libraries even without NVIDIA GPUs
- Created poor user experience
-
-## Solution
-
-### Build Process
-
-**Two separate binaries are built:**
-
-1. **voicebox-server.exe** (CPU)
-   - Built with: `pip install torch --index-url https://download.pytorch.org/whl/cpu`
-   - Size: ~500MB
-   - Works on all Windows machines
-   - Included in the installer by default
-
-2. **voicebox-server-cuda.exe** (CUDA)
-   - Built with: `pip install torch --index-url https://download.pytorch.org/whl/cu121`
-   - Size: ~3GB
-   - Requires NVIDIA GPU + drivers
-   - Uploaded as separate GitHub Release asset
-
-### User Experience
-
-**First Launch:**
-1. User installs app (~500MB download)
-2. App starts with CPU server
-3. If NVIDIA GPU detected:
-   - Show notification: "Download CUDA support for 4-5x faster inference?"
-   - User clicks "Download"
-   - Download voicebox-server-cuda.exe from GitHub (~3GB)
-   - Save to `%APPDATA%/voicebox/binaries/`
-   - Restart server with CUDA version
-
-**Settings Panel:**
- Toggle between CPU/CUDA modes
- Download CUDA if not already installed
- Show current inference backend
-
-### Build Scripts
-
-**Windows:**
-```bash
-cd backend
-
-# Build CPU only
-build_cpu.bat
-
-# Build CUDA only
-build_cuda.bat
-
-# Build both
-build_both.bat
-```
-
-**Unix (macOS/Linux):**
-```bash
-cd backend
-
-# Build CPU only
-./build_cpu.sh
-```
-
-### CI/CD Workflow
-
-**GitHub Actions (.github/workflows/release.yml):**
-
-1. Install CPU PyTorch
-2. Build CPU server → Copy to Tauri binaries
-3. Install CUDA PyTorch
-4. Build CUDA server → Save for upload
-5. Build Tauri app (bundles CPU server)
-6. Upload CUDA server as separate release asset
-
-### File Structure
-
-```
-Release Assets:
-├── Voicebox_0.1.12_x64_en-US.msi           (~500MB - includes CPU server)
-├── voicebox-server-cuda-x86_64-pc-windows-msvc.exe   (~3GB - optional download)
-└── latest.json                              (updater manifest)
-```
-
-## Implementation Details
-
-### Modified Files
-
-1. **backend/build_binary.py**
-   - Added `variant` parameter ('cpu' or 'cuda')
-   - Outputs different binary names based on variant
-
-2. **backend/build_cpu.bat** (new)
-   - Installs CPU PyTorch
-   - Builds CPU binary
-   - Restores CUDA PyTorch for dev
-
-3. **backend/build_cuda.bat** (new)
-   - Ensures CUDA PyTorch is installed
-   - Builds CUDA binary
-
-4. **.github/workflows/release.yml**
-   - Build CPU binary first (for installer)
-   - Build CUDA binary second (for upload)
-   - Upload CUDA binary as additional release asset
-   - Updated release notes to explain GPU acceleration
-
-### Future Frontend Work
-
-**TODO: Implement CUDA download in the app**
-
-Location: `tauri/src/`
-
-Features needed:
-1. GPU detection on startup
-2. Download manager for CUDA binary
-3. Server binary path switcher
-4. Settings UI for CPU/CUDA toggle
-5. Progress indicator for 3GB download
-
-API endpoints needed (already exist):
- `/health` - Shows GPU availability
- Server restart mechanism
-
-## Benefits
-
-✓ **Smaller installer**: ~500MB instead of 3GB
-✓ **Faster CI builds**: WiX can handle 500MB easily
-✓ **User choice**: CPU users don't download unnecessary files
-✓ **Better UX**: Optional performance upgrade for GPU users
-✓ **Cost savings**: Reduced bandwidth for users without GPUs
-
-## Testing
-
-**Test CPU build:**
-```bash
-cd backend
-python build_binary.py cpu
-./dist/voicebox-server.exe --version
-```
-
-**Test CUDA build:**
-```bash
-cd backend
-python build_binary.py cuda
-./dist/voicebox-server-cuda.exe --version
-```
-
-**Verify size:**
-```bash
-ls -lh backend/dist/
-# Should see:
-# voicebox-server.exe       ~500MB
-# voicebox-server-cuda.exe  ~3GB
-```
-
-**Test server startup:**
-```bash
-# CPU version
-./backend/dist/voicebox-server.exe
-# Check logs: Should show CPU inference
-
-# CUDA version (requires NVIDIA GPU)
-./backend/dist/voicebox-server-cuda.exe
-# Check logs: Should show CUDA inference
-```
@@ -1,122 +0,0 @@
-# GitHub 2GB Release Asset Limit Issue
-
-## Problem
-
-The CUDA server binary upload fails in CI with:
-```
-Error: File size (2543828017) is greater than 2 GiB
-```
-
-GitHub release assets have a hard limit of 2GB per file. Our CUDA binary is ~2.5GB, which exceeds this limit.
-
-## Background
-
-The dual-server binary system (see `dual-server-binaries.md`) creates two binaries:
- **CPU binary**: ~500MB ✅ Works fine
- **CUDA binary**: ~2.5GB ❌ Exceeds GitHub limit
-
-## Attempted Solution: Compression
-
-We're testing 7z compression with maximum settings to see if we can squeeze the CUDA binary under 2GB.
-
-### Test Script
-
-Run `backend/test_cuda_compression.py` to test compression locally:
-
-```bash
-cd backend
-python test_cuda_compression.py
-```
-
-This will:
-1. Find the CUDA binary in `dist/`
-2. Compress it with 7z (maximum compression)
-3. Report if the compressed size fits under 2GB
-
-### Expected Compression
-
-PyTorch CUDA binaries typically compress well since they contain:
- Repeated patterns in neural network weights
- Debug symbols and metadata
- Redundant CUDA libraries
-
-Estimated compression: 30-40% reduction
- Original: ~2.5GB
- Target: <2GB
- Required compression: >20%
-
-## Fallback: External Hosting
-
-If compression doesn't work, we'll need to host the CUDA binary externally:
-
-### Option 1: AWS S3
-```yaml
- name: Upload CUDA binary to S3
-  run: |
-    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
-      s3://voicebox-releases/cuda-binaries/${{ github.ref_name }}/
-```
-
-### Option 2: Azure Blob Storage
-```yaml
- name: Upload to Azure Blob
-  run: |
-    az storage blob upload \
-      --account-name voiceboxreleases \
-      --container-name cuda-binaries \
-      --file backend/cuda-release/voicebox-server-cuda-*.exe
-```
-
-### Option 3: GitHub Packages (Container Registry)
-Package as a container image, though this adds complexity for desktop app distribution.
-
-## Implementation Plan
-
-1. **Test compression locally** ← Current step
-2. **If compression works (<2GB)**:
-   - Update CI to compress before upload
-   - Update app to handle .7z downloads
-   - Add extraction step in download manager
-
-3. **If compression fails (≥2GB)**:
-   - Set up external storage (likely S3)
-   - Update CI to upload to S3
-   - Provide download URL in release notes
-   - Update app download manager to fetch from S3
-
-## CI Workflow Changes (if compression works)
-
-```yaml
- name: Compress CUDA binary (Windows only)
-  if: matrix.platform == 'windows-latest'
-  shell: bash
-  run: |
-    cd backend/cuda-release
-    7z a -t7z -m0=lzma2 -mx=9 -mfb=64 -md=32m -ms=on \
-      voicebox-server-cuda-x86_64-pc-windows-msvc.7z \
-      voicebox-server-cuda-*.exe
-
- name: Upload compressed CUDA server (Windows only)
-  if: matrix.platform == 'windows-latest'
-  uses: softprops/action-gh-release@v1
-  with:
-    files: backend/cuda-release/*.7z
-```
-
-## User Experience Impact
-
-### With Compression
- Download: `voicebox-server-cuda-*.7z` (~1.5-1.8GB)
- App extracts automatically
- One extra step but manageable
-
-### With External Hosting
- Download from S3/Azure URL
- No GitHub release asset dependency
- Potentially faster download speeds (CDN)
-
-## Status
-
-🔄 **Testing compression locally to determine viability**
-
-Results pending from local test run.
@@ -5,7 +5,7 @@ description: "Welcome to Voicebox - the open-source voice synthesis studio"

 ## What is Voicebox?

-Voicebox is a **local-first voice cloning studio** with DAW-like features for professional voice synthesis. Think of it as the **Ollama for voice** — download models, clone voices, and generate speech entirely on your machine.
+Voicebox is a **local-first voice cloning studio** with DAW-like features for professional voice synthesis. Think of it as a **local, free and open-source alternative to ElevenLabs** — download models, clone voices, and generate speech entirely on your machine.

 <Frame>
  <img src="/images/app-screenshot-1.webp" alt="Voicebox App Screenshot" />
@@ -0,0 +1,964 @@
+# TTS Provider Architecture
+
+**Status:** Planned for v0.1.13
+**Created:** 2025-01-31
+**Problem:** GitHub 2GB release limit + poor UX for frequent updates requiring 2.4GB re-downloads
+
+---
+
+## Overview
+
+Split the monolithic backend into modular components:
+
+1. **Main App** (~150-200MB): Tauri + FastAPI backend + Whisper + UI/profiles/history
+2. **TTS Providers** (downloadable plugins): Separate executables for model inference
+
+This architecture solves:
+
+- ✅ GitHub 2GB release artifact limit
+- ✅ Frequent app updates without re-downloading large python binaries
+- ✅ User choice of compute backend (CPU/GPU/Cloud)
+- ✅ External provider support (OpenAI, custom servers)
+- ✅ Future extensibility
+
+---
+
+## Architecture Diagram
+
+```
+┌─────────────────────────────────────────────────────────┐
+│  Voicebox App (Tauri + Backend)           ~150MB        │
+│  ├─ UI Layer (React)                                    │
+│  ├─ Backend (FastAPI)                                   │
+│  │  ├─ Voice Profiles                                   │
+│  │  ├─ Generation History                               │
+│  │  ├─ Audio Editing / Stories                          │
+│  │  └─ Provider Manager ◄──────────────┐                │
+│  └─ Whisper (bundled, tiny ~50MB)      │                │
+└─────────────────────────────────────────┼────────────────┘
+                                          │
+                            HTTP/IPC      │
+                                          │
+         ┌────────────────────────────────┼─────────────────┐
+         │                                │                 │
+         ▼                                ▼                 ▼
+┌─────────────────┐      ┌─────────────────┐   ┌──────────────────┐
+│ TTS Provider:   │      │ TTS Provider:   │   │ TTS Provider:    │
+│ PyTorch CPU     │      │ PyTorch CUDA    │   │ MLX (Apple)      │
+│                 │      │                 │   │                  │
+│ ~300MB          │      │ ~2.4GB          │   │ ~800MB           │
+│                 │      │                 │   │                  │
+│ Local inference │      │ GPU inference   │   │ Metal inference  │
+└─────────────────┘      └─────────────────┘   └──────────────────┘
+         │                        │                     │
+         └────────────────────────┴─────────────────────┘
+                                  │
+                    ┌─────────────▼──────────────┐
+                    │  Future Providers:         │
+                    │  • Remote Server           │
+                    │  • OpenAI API              │
+                    │  • ElevenLabs              │
+                    │  • Custom Docker Container │
+                    └────────────────────────────┘
+```
+
+---
+
+## Problem Statement
+
+### Current Architecture Issues
+
+**Monolithic Binary:**
+
+- CPU version: ~295MB
+- CUDA version: ~2.37GB
+- GitHub releases: 2GB file size limit (BLOCKED)
+- Updates require re-downloading entire binary
+- Poor UX: update app → restart → download CUDA update → restart again
+
+**User Pain Points:**
+
+1. Cannot release CUDA version on GitHub (over 2GB)
+2. Every app update forces 2.4GB re-download for GPU users
+3. No flexibility (can't use OpenAI, remote servers, etc.)
+4. Wastes bandwidth for small bug fixes
+
+---
+
+## Solution: Pluggable TTS Providers
+
+### Component Breakdown
+
+#### 1. Main App (voicebox.exe / .app / .AppImage)
+
+**Size:** ~100-150MB
+
+**Includes:**
+
+- Tauri runtime + React UI
+- FastAPI backend (pure Python, no PyTorch)
+- Whisper model (tiny, ~50MB)
+- SQLite database
+- Profile/history/audio editing logic
+- Provider management system
+
+**Does NOT include:**
+
+- PyTorch (CPU or CUDA)
+- TTS models (Qwen3-TTS)
+- Heavy ML dependencies
+
+**Updates frequently:** UI fixes, feature additions, non-ML changes
+
+---
+
+#### 2. TTS Provider: PyTorch CPU
+
+**Binary:** `tts-provider-pytorch-cpu.exe`
+**Size:** ~200MB
+
+**Includes:**
+
+- PyTorch CPU build
+- Qwen3-TTS package
+- Transformers
+- No CUDA libraries
+
+**Download source:** Cloudflare R2
+**Updates rarely:** Only when model code changes
+
+---
+
+#### 3. TTS Provider: PyTorch CUDA
+
+**Binary:** `tts-provider-pytorch-cuda.exe`
+**Size:** ~2.4GB
+
+**Includes:**
+
+- PyTorch CUDA build (cu121)
+- Qwen3-TTS package
+- CUDA runtime, cuDNN, cuBLAS
+- Transformers
+
+**Download source:** Cloudflare R2
+**Platform:** Windows + Linux (NVIDIA GPU)
+**Updates rarely:** Only when model code or CUDA version changes
+
+---
+
+#### 4. TTS Provider: MLX
+
+**Binary:** `tts-provider-mlx`
+**Size:** ~150MB
+
+**Includes:**
+
+- MLX framework
+- MLX-optimized Qwen3-TTS
+- Metal acceleration
+
+**Platform:** macOS only (Apple Silicon)
+**Download source:** Cloudflare R2
+
+---
+
+#### 5. TTS Provider: Remote
+
+**Binary:** None (built-in config)
+**Size:** 0MB
+
+**How it works:**
+
+- User provides URL to their own TTS server
+- Backend proxies requests to that server
+- Implements API spec from `EXTERNAL_PROVIDERS.md`
+
+**Use cases:**
+
+- AMD GPU users running their own server
+- Team deployments with shared GPU server
+- Cloud hosting (Modal, RunPod, Replicate)
+
+---
+
+#### 6. TTS Provider: OpenAI
+
+**Binary:** None (API wrapper)
+**Size:** 0MB
+
+**How it works:**
+
+- User provides OpenAI API key
+- Backend wraps OpenAI Audio API
+- Voice profiles map to OpenAI voices
+
+**Benefits:**
+
+- Zero local compute
+- Pay-per-use
+- Instant setup
+
+---
+
+## Communication Protocol
+
+### Provider API Specification
+
+All TTS providers must implement these endpoints:
+
+#### POST /tts/generate
+
+Generate speech from text.
+
+**Request:**
+
+```json
+{
+	"text": "Hello world!",
+	"voice_prompt": {
+		/* voice prompt object */
+	},
+	"language": "en",
+	"seed": 12345,
+	"model_size": "1.7B"
+}
+```
+
+**Response:**
+
+```json
+{
+	"audio": "base64-encoded-audio",
+	"sample_rate": 24000,
+	"duration": 2.5
+}
+```
+
+#### POST /tts/create_voice_prompt
+
+Create voice prompt from reference audio.
+
+**Request:** (multipart/form-data)
+
+- `audio`: Audio file
+- `reference_text`: Transcript
+
+**Response:**
+
+```json
+{
+	"voice_prompt": {
+		/* serialized prompt */
+	}
+}
+```
+
+#### GET /tts/health
+
+Health check.
+
+**Response:**
+
+```json
+{
+	"status": "healthy",
+	"provider": "pytorch-cuda",
+	"version": "1.0.0",
+	"model": "Qwen3-TTS-12Hz-1.7B-Base",
+	"device": "cuda:0"
+}
+```
+
+#### GET /tts/status
+
+Model status.
+
+**Response:**
+
+```json
+{
+	"model_loaded": true,
+	"model_size": "1.7B",
+	"available_sizes": ["0.6B", "1.7B"],
+	"gpu_available": true,
+	"vram_used_mb": 1234
+}
+```
+
+---
+
+## Backend Implementation
+
+### Provider Manager
+
+**File:** `backend/providers/__init__.py`
+
+```python
+class ProviderManager:
+    """Manages TTS provider lifecycle."""
+
+    def __init__(self):
+        self.active_provider: Optional[Provider] = None
+        self.config = load_provider_config()
+
+    async def start_provider(self, provider_type: str) -> str:
+        """Start a TTS provider process."""
+        if provider_type == "pytorch-cpu":
+            return await self._start_local_provider("tts-provider-pytorch-cpu.exe")
+        elif provider_type == "pytorch-cuda":
+            return await self._start_local_provider("tts-provider-pytorch-cuda.exe")
+        elif provider_type == "mlx":
+            return await self._start_local_provider("tts-provider-mlx")
+        elif provider_type == "remote":
+            return self.config["remote_url"]
+        elif provider_type == "openai":
+            return None  # No subprocess, API wrapper
+
+    async def _start_local_provider(self, binary_name: str) -> str:
+        """Start local provider subprocess."""
+        provider_path = get_provider_binary_path(binary_name)
+
+        if not provider_path.exists():
+            raise ProviderNotInstalledException(binary_name)
+
+        # Start subprocess on random port
+        port = get_free_port()
+        process = subprocess.Popen([
+            str(provider_path),
+            "--port", str(port),
+            "--data-dir", str(config.get_data_dir())
+        ])
+
+        # Wait for provider to be ready
+        await wait_for_provider_health(f"http://localhost:{port}")
+
+        self.active_provider = Provider(process, port)
+        return f"http://localhost:{port}"
+
+    async def stop_provider(self):
+        """Stop active provider."""
+        if self.active_provider:
+            self.active_provider.process.terminate()
+            self.active_provider = None
+```
+
+---
+
+### Provider Abstraction
+
+**File:** `backend/providers/base.py`
+
+```python
+class TTSProvider(ABC):
+    """Abstract base for TTS providers."""
+
+    @abstractmethod
+    async def generate(
+        self,
+        text: str,
+        voice_prompt: dict,
+        language: str,
+        seed: Optional[int]
+    ) -> tuple[np.ndarray, int]:
+        """Generate speech audio."""
+        pass
+
+    @abstractmethod
+    async def create_voice_prompt(
+        self,
+        audio_path: str,
+        reference_text: str
+    ) -> dict:
+        """Create voice prompt from reference audio."""
+        pass
+```
+
+**File:** `backend/providers/local.py`
+
+```python
+class LocalProvider(TTSProvider):
+    """Provider that communicates with local subprocess via HTTP."""
+
+    def __init__(self, base_url: str):
+        self.base_url = base_url
+        self.client = httpx.AsyncClient()
+
+    async def generate(self, text, voice_prompt, language, seed):
+        response = await self.client.post(
+            f"{self.base_url}/tts/generate",
+            json={
+                "text": text,
+                "voice_prompt": voice_prompt,
+                "language": language,
+                "seed": seed
+            }
+        )
+        data = response.json()
+        audio = np.frombuffer(base64.b64decode(data["audio"]), dtype=np.float32)
+        return audio, data["sample_rate"]
+```
+
+**File:** `backend/providers/openai.py`
+
+```python
+class OpenAIProvider(TTSProvider):
+    """Provider that wraps OpenAI Audio API."""
+
+    def __init__(self, api_key: str):
+        self.client = OpenAI(api_key=api_key)
+
+    async def generate(self, text, voice_prompt, language, seed):
+        # Map voice_prompt to OpenAI voice name
+        voice = map_profile_to_openai_voice(voice_prompt)
+
+        response = await self.client.audio.speech.create(
+            model="tts-1",
+            voice=voice,
+            input=text
+        )
+
+        # Convert to numpy array
+        audio_data = response.content
+        audio, sr = load_audio_from_bytes(audio_data)
+        return audio, sr
+```
+
+---
+
+## Provider Installation
+
+### Download Manager
+
+**File:** `backend/providers/installer.py`
+
+```python
+class ProviderInstaller:
+    """Handles provider download and installation."""
+
+    async def download_provider(self, provider_type: str):
+        """Download provider binary from R2."""
+
+        binary_name = {
+            "pytorch-cpu": "tts-provider-pytorch-cpu.exe",
+            "pytorch-cuda": "tts-provider-pytorch-cuda.exe",
+            "mlx": "tts-provider-mlx"
+        }[provider_type]
+
+        download_url = f"https://downloads.voicebox.sh/providers/v{PROVIDER_VERSION}/{binary_name}"
+
+        # Download with progress tracking (reuse existing SSE system)
+        await download_with_progress(
+            url=download_url,
+            destination=get_provider_install_path(binary_name),
+            progress_key=f"provider-{provider_type}"
+        )
+```
+
+**Provider Storage Location:**
+
+- Windows: `%APPDATA%/voicebox/providers/`
+- macOS: `~/Library/Application Support/voicebox/providers/`
+- Linux: `~/.local/share/voicebox/providers/`
+
+---
+
+## Frontend Implementation
+
+### Provider Settings UI
+
+**Component:** `app/src/components/ServerSettings/ProviderSettings.tsx`
+
+```tsx
+export function ProviderSettings() {
+	const [selectedProvider, setSelectedProvider] =
+		useState<ProviderType>("auto");
+	const {data: installedProviders} = useQuery({
+		queryKey: ["providers", "installed"],
+		queryFn: () => apiClient.getInstalledProviders(),
+	});
+
+	return (
+		<Card>
+			<CardHeader>
+				<CardTitle>TTS Provider</CardTitle>
+				<CardDescription>Choose how Voicebox generates speech</CardDescription>
+			</CardHeader>
+			<CardContent>
+				<RadioGroup
+					value={selectedProvider}
+					onValueChange={setSelectedProvider}
+				>
+					{/* Auto-detect */}
+					<div className="flex items-center space-x-2">
+						<RadioGroupItem value="auto" id="auto" />
+						<Label htmlFor="auto">
+							<div className="font-medium">Auto-detect (Recommended)</div>
+							<div className="text-sm text-muted-foreground">
+								Automatically choose the best available provider
+							</div>
+						</Label>
+					</div>
+
+					{/* PyTorch CUDA */}
+					<div className="flex items-center justify-between">
+						<div className="flex items-center space-x-2">
+							<RadioGroupItem
+								value="pytorch-cuda"
+								id="cuda"
+								disabled={!gpuAvailable}
+							/>
+							<Label htmlFor="cuda">
+								<div className="font-medium">PyTorch CUDA (NVIDIA GPU)</div>
+								<div className="text-sm text-muted-foreground">
+									4-5x faster inference on NVIDIA GPUs
+								</div>
+							</Label>
+						</div>
+						{!installedProviders?.includes("pytorch-cuda") && gpuAvailable && (
+							<Button
+								onClick={() => downloadProvider("pytorch-cuda")}
+								size="sm"
+							>
+								Download (2.4GB)
+							</Button>
+						)}
+					</div>
+
+					{/* PyTorch CPU */}
+					<div className="flex items-center justify-between">
+						<div className="flex items-center space-x-2">
+							<RadioGroupItem value="pytorch-cpu" id="cpu" />
+							<Label htmlFor="cpu">
+								<div className="font-medium">PyTorch CPU</div>
+								<div className="text-sm text-muted-foreground">
+									Works on any system, slower inference
+								</div>
+							</Label>
+						</div>
+						{!installedProviders?.includes("pytorch-cpu") && (
+							<Button onClick={() => downloadProvider("pytorch-cpu")} size="sm">
+								Download (300MB)
+							</Button>
+						)}
+					</div>
+
+					{/* MLX (macOS only) */}
+					{isMacOS && (
+						<div className="flex items-center justify-between">
+							<div className="flex items-center space-x-2">
+								<RadioGroupItem value="mlx" id="mlx" />
+								<Label htmlFor="mlx">
+									<div className="font-medium">MLX (Apple Silicon)</div>
+									<div className="text-sm text-muted-foreground">
+										Optimized for M1/M2/M3 chips
+									</div>
+								</Label>
+							</div>
+							{!installedProviders?.includes("mlx") && (
+								<Button onClick={() => downloadProvider("mlx")} size="sm">
+									Download (800MB)
+								</Button>
+							)}
+						</div>
+					)}
+
+					{/* Remote */}
+					<div className="space-y-2">
+						<div className="flex items-center space-x-2">
+							<RadioGroupItem value="remote" id="remote" />
+							<Label htmlFor="remote">
+								<div className="font-medium">Remote Server</div>
+								<div className="text-sm text-muted-foreground">
+									Connect to your own TTS server
+								</div>
+							</Label>
+						</div>
+						{selectedProvider === "remote" && (
+							<Input placeholder="http://your-server:8000" className="ml-6" />
+						)}
+					</div>
+
+					{/* OpenAI */}
+					<div className="space-y-2">
+						<div className="flex items-center space-x-2">
+							<RadioGroupItem value="openai" id="openai" />
+							<Label htmlFor="openai">
+								<div className="font-medium">OpenAI API</div>
+								<div className="text-sm text-muted-foreground">
+									Use OpenAI's TTS API (requires API key)
+								</div>
+							</Label>
+						</div>
+						{selectedProvider === "openai" && (
+							<Input type="password" placeholder="sk-..." className="ml-6" />
+						)}
+					</div>
+				</RadioGroup>
+			</CardContent>
+		</Card>
+	);
+}
+```
+
+---
+
+## File Structure
+
+```
+voicebox/
+├── backend/
+│   ├── main.py                    # Main FastAPI app (no TTS code)
+│   ├── providers/
+│   │   ├── __init__.py            # ProviderManager
+│   │   ├── base.py                # TTSProvider ABC
+│   │   ├── local.py               # LocalProvider (subprocess)
+│   │   ├── remote.py              # RemoteProvider (HTTP)
+│   │   ├── openai.py              # OpenAIProvider (API wrapper)
+│   │   └── installer.py           # Provider download logic
+│   ├── profiles.py                # Voice profile management
+│   ├── history.py                 # Generation history
+│   ├── transcribe.py              # Whisper (still bundled)
+│   └── ... (other backend modules)
+│
+├── providers/
+│   ├── pytorch-cpu/
+│   │   ├── main.py                # FastAPI server for TTS
+│   │   ├── tts_backend.py         # PyTorch TTS logic
+│   │   ├── requirements.txt       # torch (CPU), qwen-tts, transformers
+│   │   └── build.spec             # PyInstaller spec
+│   │
+│   ├── pytorch-cuda/
+│   │   ├── main.py                # FastAPI server for TTS
+│   │   ├── tts_backend.py         # PyTorch TTS logic
+│   │   ├── requirements.txt       # torch+cu121, qwen-tts, transformers
+│   │   └── build.spec             # PyInstaller spec
+│   │
+│   └── mlx/
+│       ├── main.py                # FastAPI server for TTS
+│       ├── mlx_backend.py         # MLX TTS logic
+│       ├── requirements.txt       # mlx, qwen-tts-mlx
+│       └── build.spec             # PyInstaller spec
+│
+├── app/                           # Frontend (Tauri + React)
+│   └── src/
+│       └── components/
+│           └── ServerSettings/
+│               └── ProviderSettings.tsx
+│
+└── tauri/
+    └── src-tauri/
+        └── tauri.conf.json        # No externalBin for providers
+```
+
+---
+
+## Migration Path
+
+### Phase 1: Refactor Backend (No User Changes)
+
+**Goal:** Abstract TTS behind provider interface
+
+1. Create `backend/providers/` module structure
+2. Implement `TTSProvider` abstract base class
+3. Create `LocalProvider` wrapper for current PyTorch code
+4. Modify `backend/tts.py` to use provider abstraction
+5. Keep PyTorch bundled in main app
+
+**Result:** Code is prepared, but user experience unchanged
+
+---
+
+### Phase 2: Build Provider Binaries
+
+**Goal:** Create standalone TTS provider executables
+
+1. Create separate PyInstaller specs for each provider
+2. Build provider executables:
+   - `tts-provider-pytorch-cpu.exe` (~300MB)
+   - `tts-provider-pytorch-cuda.exe` (~2.4GB)
+   - `tts-provider-mlx` (~800MB, macOS)
+3. Test subprocess communication
+4. Upload providers to Cloudflare R2
+
+**Result:** Provider binaries exist but aren't used yet
+
+---
+
+### Phase 3: Remove PyTorch from Main App
+
+**Goal:** Split main app from providers
+
+1. Exclude PyTorch/Qwen3-TTS from main app PyInstaller spec
+2. Main app now requires provider download
+3. Update GitHub CI to build multiple artifacts:
+   - `voicebox-{version}-{platform}.exe` (~150MB)
+   - `tts-provider-pytorch-cpu-{version}.exe`
+   - `tts-provider-pytorch-cuda-{version}.exe`
+   - `tts-provider-mlx-{version}` (macOS)
+
+**Result:** Main app is small, providers downloaded separately
+
+---
+
+### Phase 4: Add Provider UI
+
+**Goal:** User-facing provider management
+
+1. Create Provider Settings page
+2. Implement provider download UI
+3. Add provider status indicators
+4. Show active provider in UI
+
+**Result:** Users can choose and download providers
+
+---
+
+### Phase 5: External Providers
+
+**Goal:** Enable remote and cloud providers
+
+1. Implement `RemoteProvider` (HTTP client)
+2. Implement `OpenAIProvider` (API wrapper)
+3. Add provider configuration UI (URLs, API keys)
+4. Document external provider API spec
+
+**Result:** Full provider ecosystem
+
+---
+
+## Provider Versioning
+
+### Independent Versioning
+
+Providers have their own version numbers, independent of the main app:
+
+- **App version:** `v0.2.0` (frequent updates)
+- **Provider version:** `v1.0.0` (rare updates)
+
+### Compatibility Matrix
+
+**Example:**
+
+| App Version | Min Provider Version | Max Provider Version |
+| ----------- | -------------------- | -------------------- |
+| v0.2.0      | v1.0.0               | v1.x.x               |
+| v0.3.0      | v1.0.0               | v1.x.x               |
+| v0.4.0      | v1.2.0               | v1.x.x               |
+| v1.0.0      | v2.0.0               | v2.x.x               |
+
+**Backend checks compatibility:**
+
+```python
+async def check_provider_compatibility(provider_version: str) -> bool:
+    """Check if provider version is compatible with current app."""
+    min_version = "1.0.0"
+    max_version = "1.999.999"
+    return min_version <= provider_version < max_version
+```
+
+**UI shows warning if incompatible:**
+
+```
+⚠️ Provider version 0.9.0 is outdated. Update to v1.0.0+
+```
+
+---
+
+## User Flows
+
+### First-Time Setup
+
+1. User downloads and installs Voicebox (~150MB)
+2. App launches → detects no TTS provider installed
+3. Shows setup wizard:
+
+   ```
+   Choose your TTS provider:
+
+   [ ] PyTorch CUDA (2.4GB)    [Download]
+       ✓ Fastest on NVIDIA GPUs
+       ✗ Requires NVIDIA GPU
+
+   [●] PyTorch CPU (300MB)     [Download]
+       ✓ Works on any system
+       ✗ Slower inference
+
+   [ ] MLX (800MB)             [Download]
+       ✓ Fast on Apple Silicon
+       ✗ macOS only (M1/M2/M3)
+
+   [ ] Remote Server
+       URL: ___________________
+
+   [ ] OpenAI API
+       API Key: ________________
+   ```
+
+4. User selects provider → downloads with progress bar
+5. Provider installs to AppData/Application Support
+6. App starts provider → ready to use
+
+---
+
+### App Update Flow (No Provider Change)
+
+**Scenario:** Bug fix in UI, no backend changes
+
+1. User gets update notification: "Voicebox v0.2.1 available"
+2. Downloads update (~150MB, not 2.4GB!)
+3. Installs and restarts
+4. **Provider stays the same** (no re-download needed)
+5. App starts using existing provider
+
+**User experience:** Fast updates, no multi-GB downloads
+
+---
+
+### Provider Update Flow
+
+**Scenario:** New Qwen3-TTS model version released
+
+1. User opens Settings → Provider tab
+2. Sees notification: "Provider update available (v1.1.0)"
+3. Clicks "Update Provider"
+4. Downloads new provider binary
+5. Old provider binary is replaced
+6. Restart app to use new provider
+
+**Frequency:** Rare (only when TTS model/backend changes)
+
+---
+
+### Switching Providers
+
+**Scenario:** User upgrades to NVIDIA GPU
+
+1. User goes to Settings → Provider
+2. Selects "PyTorch CUDA"
+3. Clicks "Download" → downloads 2.4GB
+4. Download completes → restarts app
+5. App now uses CUDA provider
+
+---
+
+## Benefits
+
+| Benefit                       | Details                                                   |
+| ----------------------------- | --------------------------------------------------------- |
+| **GitHub Releases Work**      | Main app ~150MB << 2GB limit                              |
+| **Fast Updates**              | UI/feature updates don't require re-downloading providers |
+| **User Choice**               | CPU, CUDA, MLX, OpenAI, remote server                     |
+| **External Provider Support** | Users can run their own TTS servers                       |
+| **Bandwidth Savings**         | Only download provider once, app updates are small        |
+| **Future-Proof**              | Easy to add new providers (ElevenLabs, custom models)     |
+| **Team Deployments**          | Multiple users share one remote provider                  |
+| **Cloud-Ready**               | Works with Modal, Replicate, RunPod, etc.                 |
+
+---
+
+## Open Questions
+
+### 1. Provider Versioning
+
+**Question:** Should providers have independent versions or match app version?
+
+**Options:**
+
+- A. Independent (providers: v1.x, app: v0.2.x)
+- B. Matched (both use v0.2.x)
+
+**Recommendation:** Independent versioning with compatibility matrix
+
+---
+
+### 2. Auto-Update Providers
+
+**Question:** Should providers auto-update separately from app?
+
+**Options:**
+
+- A. Manual updates only (user clicks "Update Provider")
+- B. Optional auto-update (user can enable)
+- C. Always auto-update
+
+**Recommendation:** Optional auto-update (default off)
+
+---
+
+### 3. Provider Discovery
+
+**Question:** How does app find installed providers?
+
+**Options:**
+
+- A. Check standard paths in AppData/Application Support
+- B. Registry (Windows) / plist (macOS)
+- C. Config file with provider locations
+
+**Recommendation:** Standard paths + config fallback
+
+---
+
+### 4. Fallback Behavior
+
+**Question:** What if no provider is installed?
+
+**Options:**
+
+- A. Show setup wizard on first launch
+- B. Block app until provider installed
+- C. Allow app to run in "demo mode" (transcription only)
+
+**Recommendation:** Setup wizard on first launch
+
+---
+
+### 5. Provider Auto-Start
+
+**Question:** Should provider start automatically with app?
+
+**Options:**
+
+- A. Always start selected provider on app launch
+- B. Start on-demand (when user generates speech)
+- C. User preference
+
+**Recommendation:** Auto-start (configurable in settings)
+
+---
+
+## Future Enhancements
+
+- [ ] **Provider Marketplace:** Built-in directory of community providers
+- [ ] **Multi-Provider Support:** Use different providers per voice/language
+- [ ] **Provider Health Monitoring:** Automatic failover if provider crashes
+- [ ] **Cost Tracking:** Monitor API usage for OpenAI/cloud providers
+- [ ] **Performance Metrics:** Latency, throughput, VRAM usage dashboards
+- [ ] **Docker Providers:** Run providers in Docker containers
+- [ ] **Provider Plugins:** Load custom providers from user scripts
+
+---
+
+## Related Documents
+
+- [EXTERNAL_PROVIDERS.md](./EXTERNAL_PROVIDERS.md) - External provider support plan
+- [OPENAI_SUPPORT.md](./OPENAI_SUPPORT.md) - OpenAI API compatibility
+- [github-2gb-limit-issue.md](../github-2gb-limit-issue.md) - Original problem
+- [r2-setup.md](../r2-setup.md) - Cloudflare R2 configuration
+
+---
+
+## Contributing
+
+If you want to build a custom TTS provider:
+
+1. Implement the provider API spec (see above)
+2. Test with Voicebox locally
+3. Package as executable (PyInstaller, Docker, etc.)
+4. Share in GitHub Discussions
+
+**Questions?**
+
+- GitHub Issues: [voicebox/issues](https://github.com/jamiepine/voicebox/issues)
+- Discord: Coming soon
@@ -1,274 +0,0 @@
-# Cloudflare R2 Setup Guide
-
-## Overview
-
-The CUDA binary (2.4GB) is hosted on Cloudflare R2 at `downloads.voicebox.sh` instead of GitHub Releases (which has a 2GB limit).
-
-## R2 Bucket Configuration
-
-✅ **Completed:**
- Bucket created: `voicebox`
- Custom domain configured: `downloads.voicebox.sh`
-
-## GitHub Secrets Required
-
-Add these secrets to your GitHub repository:
-
-### 1. R2_ACCESS_KEY_ID
-
-Your Cloudflare R2 API Access Key ID
-
-**How to get it:**
-1. Go to Cloudflare Dashboard → R2
-2. Click "Manage R2 API Tokens"
-3. Create API Token with "Object Read & Write" permissions
-4. Copy the "Access Key ID"
-
-**Add to GitHub:**
-```
-Repository Settings → Secrets and variables → Actions → New repository secret
-Name: R2_ACCESS_KEY_ID
-Value: <your-access-key-id>
-```
-
-### 2. R2_SECRET_ACCESS_KEY
-
-Your Cloudflare R2 Secret Access Key
-
-**How to get it:**
- Same process as above
- Copy the "Secret Access Key" (shown only once!)
- Store it securely
-
-**Add to GitHub:**
-```
-Name: R2_SECRET_ACCESS_KEY
-Value: <your-secret-access-key>
-```
-
-### 3. R2_ENDPOINT
-
-Your Cloudflare R2 endpoint URL
-
-**Format:**
-```
-https://<account-id>.r2.cloudflarestorage.com
-```
-
-**How to find your account ID:**
- Cloudflare Dashboard → R2
- Look at the URL or bucket settings
- Should be a string of letters/numbers
-
-**Add to GitHub:**
-```
-Name: R2_ENDPOINT
-Value: https://<your-account-id>.r2.cloudflarestorage.com
-```
-
-## Bucket Structure
-
-After CI uploads, the bucket will have this structure:
-
-```
-voicebox/
-└── cuda/
-    ├── v0.1.12/
-    │   └── voicebox-server-cuda-x86_64-pc-windows-msvc.exe
-    ├── v0.1.13/
-    │   └── voicebox-server-cuda-x86_64-pc-windows-msvc.exe
-    └── v0.2.0/
-        └── voicebox-server-cuda-x86_64-pc-windows-msvc.exe
-```
-
-## Public Access
-
-Files are uploaded with `--acl public-read`, making them accessible at:
-
-```
-https://downloads.voicebox.sh/cuda/v{VERSION}/voicebox-server-cuda-x86_64-pc-windows-msvc.exe
-```
-
-**Example:**
-```
-https://downloads.voicebox.sh/cuda/v0.1.12/voicebox-server-cuda-x86_64-pc-windows-msvc.exe
-```
-
-## Testing the Setup
-
-### Local Test Upload
-
-Before running the CI, test uploading locally:
-
-```bash
-# Set environment variables
-export AWS_ACCESS_KEY_ID="your-r2-access-key-id"
-export AWS_SECRET_ACCESS_KEY="your-r2-secret-access-key"
-export R2_ENDPOINT="https://your-account-id.r2.cloudflarestorage.com"
-
-# Install AWS CLI
-pip install awscli
-
-# Test upload (use a small test file first)
-echo "test" > test.txt
-aws s3 cp test.txt \
-  s3://voicebox/test/test.txt \
-  --endpoint-url $R2_ENDPOINT \
-  --acl public-read
-
-# Verify it's accessible
-curl https://downloads.voicebox.sh/test/test.txt
-
-# If successful, try the actual CUDA binary
-aws s3 cp backend/dist/voicebox-server-cuda.exe \
-  s3://voicebox/cuda/v0.1.12-test/voicebox-server-cuda-x86_64-pc-windows-msvc.exe \
-  --endpoint-url $R2_ENDPOINT \
-  --acl public-read
-```
-
-### Verify Upload
-
-Check if the file is accessible:
-
-```bash
-curl -I https://downloads.voicebox.sh/cuda/v0.1.12-test/voicebox-server-cuda-x86_64-pc-windows-msvc.exe
-```
-
-Should return:
-```
-HTTP/2 200
-content-length: 2545086396
-content-type: application/x-msdownload
-...
-```
-
-## CI Workflow
-
-The workflow now:
-
-1. **Builds CPU binary** → Includes in installer
-2. **Builds CUDA binary** → Uploads to R2
-3. **Release notes** → Include R2 download link
-
-### CI Steps (Windows)
-
-```yaml
- name: Build CUDA Python server (Windows only)
-  # Builds the CUDA binary
-
- name: Upload CUDA server to Cloudflare R2 (Windows only)
-  env:
-    AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
-    AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
-    R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
-  run: |
-    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
-      s3://voicebox/cuda/${VERSION}/... \
-      --endpoint-url $R2_ENDPOINT \
-      --acl public-read
-```
-
-## Cost Tracking
-
-Monitor your R2 usage:
-
-**Cloudflare Dashboard → R2 → voicebox → Metrics**
-
-Expected costs (per month):
- Storage: 2.4GB × $0.015/GB = **$0.036**
- Egress: **$0.00** (free!)
- Class A ops: ~100 × $4.50/million = **$0.00**
-
-**Total: ~$0.04/month** (essentially free!)
-
-## Troubleshooting
-
-### Upload fails: "Access Denied"
-
-**Solution:** Check API token permissions
- Must have "Object Read & Write" on the bucket
- Regenerate token if needed
-
-### File not accessible at downloads.voicebox.sh
-
-**Solution:** Check custom domain configuration
- R2 Dashboard → Bucket → Settings → Custom Domains
- Ensure `downloads.voicebox.sh` is properly configured
- DNS may take time to propagate
-
-### "endpoint-url" not recognized
-
-**Solution:** Make sure AWS CLI is updated
-```bash
-pip install --upgrade awscli
-```
-
-### File uploaded but wrong permissions
-
-**Solution:** Re-upload with `--acl public-read`
-```bash
-aws s3 cp ... --acl public-read
-```
-
-Or set bucket default permissions in R2 Dashboard.
-
-## Security Notes
-
-### API Token Permissions
-
-✅ **Recommended:**
- Object Read & Write only
- No admin permissions needed
- Scoped to `voicebox` bucket only
-
-❌ **Avoid:**
- Account-wide permissions
- Account admin access
- Worker edit permissions
-
-### Secret Rotation
-
-Rotate API tokens every 6-12 months:
-1. Create new API token
-2. Update GitHub secrets
-3. Verify CI still works
-4. Delete old token
-
-## Maintenance
-
-### Cleaning Old Versions
-
-Optional: Delete old CUDA binaries to save storage costs
-
-```bash
-# List all versions
-aws s3 ls s3://voicebox/cuda/ \
-  --endpoint-url $R2_ENDPOINT
-
-# Delete old version
-aws s3 rm s3://voicebox/cuda/v0.1.0/ \
-  --recursive \
-  --endpoint-url $R2_ENDPOINT
-```
-
-### Monitoring
-
-Set up Cloudflare notifications:
- Storage approaching limits
- Unusual traffic patterns
- High operation counts
-
-## Next Steps
-
-1. ✅ Bucket configured
-2. ⏳ Add GitHub secrets (R2_ACCESS_KEY_ID, R2_SECRET_ACCESS_KEY, R2_ENDPOINT)
-3. ⏳ Test local upload
-4. ⏳ Push branch and create test release
-5. ⏳ Verify CUDA binary accessible from downloads.voicebox.sh
-6. ⏳ Implement frontend download manager
-
---
-
-**Status**: Ready for testing
-**Cost**: ~$0.04/month
-**Bandwidth**: Free (unlimited)
@@ -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",
@@ -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}
@@ -239,8 +239,9 @@ export default function Home() {
          <div className="space-y-6 text-lg text-foreground/80 text-center">
            <p>
              Voicebox is a <strong>local-first voice cloning studio</strong> with DAW-like features
-              for professional voice synthesis. Think of it as the <strong>Ollama for voice</strong>{' '}
-              — download models, clone voices, and generate speech entirely on your machine.
+              for professional voice synthesis. Think of it as a{' '}
+              <strong>local, free and open-source alternative to ElevenLabs</strong> — download
+              models, clone voices, and generate speech entirely on your machine.
            </p>
            <p>
              Unlike cloud services that lock your voice data behind subscriptions, Voicebox gives
@@ -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>
+  );
+}
@@ -1,6 +1,6 @@
 {
  "name": "voicebox",
-  "version": "0.1.12",
+  "version": "0.1.13",
  "private": true,
  "workspaces": [
    "app",
@@ -0,0 +1,9 @@
+uvicorn
+fastapi
+sqlalchemy
+torch
+torchvision
+soundfile
+librosa
+python-multipart
+huggingface_hub
@@ -1,7 +1,7 @@
 {
  "name": "@voicebox/tauri",
  "private": true,
-  "version": "0.1.12",
+  "version": "0.1.13",
  "type": "module",
  "scripts": {
    "dev": "vite",
@@ -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 = ""
@@ -0,0 +1,16 @@
+use crate::audio_capture::AudioCaptureState;
+
+pub async fn start_capture(
+    state: &AudioCaptureState,
+    max_duration_secs: u32,
+) -> Result<(), String> {
+    todo!("implement Linux audio capture")
+}
+
+pub async fn stop_capture(state: &AudioCaptureState) -> Result<String, String> {
+    todo!("implement Linux audio capture stop")
+}
+
+pub fn is_supported() -> bool {
+    false
+}
@@ -2,11 +2,15 @@
 mod macos;
 #[cfg(target_os = "windows")]
 mod windows;
+#[cfg(target_os = "linux")]
+mod linux;

 #[cfg(target_os = "macos")]
 pub use macos::*;
 #[cfg(target_os = "windows")]
 pub use windows::*;
+#[cfg(target_os = "linux")]
+pub use linux::*;

 use std::sync::{Arc, Mutex};

@@ -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",
@@ -2,29 +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) {
-        const { writeBinaryFile } = await import('@tauri-apps/plugin-fs');
-        const arrayBuffer = await blob.arrayBuffer();
-        await writeBinaryFile(filePath, 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 filePath = await save({
+      defaultPath: filename,
+      filters: filters || [],
+    });
+
+    if (!filePath) return; // User cancelled the dialog
+
+    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));
  },
 };
@@ -35,15 +35,5 @@ export default defineConfig({
    minify: !process.env.TAURI_DEBUG,
    sourcemap: !!process.env.TAURI_DEBUG,
    outDir: 'dist',
-    rollupOptions: {
-      external: [
-        '@tauri-apps/api',
-        '@tauri-apps/plugin-dialog',
-        '@tauri-apps/plugin-fs',
-        '@tauri-apps/plugin-process',
-        '@tauri-apps/plugin-shell',
-        '@tauri-apps/plugin-updater',
-      ],
-    },
  },
 });
@@ -1,77 +0,0 @@
-"""Test CUDA detection in voicebox backend"""
-import sys
-import torch
-
-print("=" * 60)
-print("PyTorch CUDA Detection Test")
-print("=" * 60)
-
-# Basic torch info
-print(f"\nPyTorch version: {torch.__version__}")
-print(f"CUDA available: {torch.cuda.is_available()}")
-
-if torch.cuda.is_available():
-    print(f"CUDA version: {torch.version.cuda}")
-    print(f"GPU count: {torch.cuda.device_count()}")
-    print(f"Current GPU: {torch.cuda.current_device()}")
-    print(f"GPU name: {torch.cuda.get_device_name(0)}")
-    print(f"GPU memory: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f} GB")
-else:
-    print("\nNo CUDA available - would run on CPU")
-
-# Test backend device selection
-print("\n" + "=" * 60)
-print("Backend Device Selection")
-print("=" * 60)
-
-# Simulate the _get_device method from pytorch_backend.py
-def _get_device() -> str:
-    """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"
-    return "cpu"
-
-selected_device = _get_device()
-print(f"\nSelected device: {selected_device}")
-print(f"Would use dtype: {'torch.bfloat16' if selected_device != 'cpu' else 'torch.float32'}")
-
-# Test actual tensor creation on device
-print("\n" + "=" * 60)
-print("Testing Tensor Creation on Device")
-print("=" * 60)
-
-try:
-    test_tensor = torch.randn(1000, 1000).to(selected_device)
-    print(f"\n[OK] Successfully created tensor on {selected_device}")
-    print(f"  Tensor device: {test_tensor.device}")
-    print(f"  Tensor dtype: {test_tensor.dtype}")
-
-    # Test computation
-    result = test_tensor @ test_tensor.T
-    print(f"[OK] Successfully performed computation on {selected_device}")
-
-    if selected_device == "cuda":
-        print(f"\nCUDA memory allocated: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
-        print(f"CUDA memory reserved: {torch.cuda.memory_reserved() / 1024**2:.2f} MB")
-
-except Exception as e:
-    print(f"\n[ERROR] {e}")
-
-print("\n" + "=" * 60)
-print("Summary")
-print("=" * 60)
-
-if selected_device == "cuda":
-    print("\n[SUCCESS] CUDA IS WORKING!")
-    print("  The backend will use your NVIDIA GPU for inference")
-    print(f"  GPU: {torch.cuda.get_device_name(0)}")
-    print(f"  This will be significantly faster than CPU")
-else:
-    print("\n[FAIL] CUDA is not available")
-    print("  The backend will use CPU for inference")
-    print("  This will be slower than GPU")
-
-print("\n" + "=" * 60)
@@ -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",
@@ -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'),
Author	SHA1	Message	Date
Jamie Pine	38bf96ff20	fix: pin numba for release CI wheel compatibility	2026-02-23 12:18:34 -08:00
Jamie Pine	0b14cb1b2c	Bump version: 0.1.12 → 0.1.13	2026-02-23 11:24:46 -08:00
Jamie Pine	4d24e69012	docs: point download links to latest release	2026-02-23 11:23:30 -08:00
Jamie PineandGitHub	90436e428d	Merge pull request #77 from ManuLG/fix/broken-confirmation-modals fix: await for confirmation before deleting voices and channels	2026-02-23 11:13:12 -08:00
Jamie PineandGitHub	e4bb288904	Merge pull request #93 from iJaack/fix/mlx-apple-silicon-binary fix(mlx): bundle native libs and broaden error handling for Apple Silicon	2026-02-23 11:12:34 -08:00
Jamie PineandGitHub	6f8bc7f23b	Merge pull request #95 from CelebrityPunks/fix/model-size-selection-ignored Fix: selecting 0.6B model still downloads and uses 1.7B	2026-02-23 11:12:12 -08:00
Jamie PineandGitHub	baca111d50	Merge branch 'main' into fix/model-size-selection-ignored	2026-02-23 11:12:05 -08:00
Jamie PineandGitHub	cc298fe6d8	Merge pull request #79 from martyniukyurii/fix/unicode-content-disposition fix: handle non-ASCII filenames in Content-Disposition headers	2026-02-23 11:09:36 -08:00
Jamie PineandGitHub	46b8f6b882	Merge pull request #78 from tomasmach/fix/getUserMedia-undefined-check fix: guard getUserMedia call against undefined mediaDevices in non-secure contexts	2026-02-23 11:09:23 -08:00
Jamie PineandGitHub	162cf4fb84	Merge pull request #122 from white1107/fix/web-tailwind-plugin fix(web): add @tailwindcss/vite plugin to web config	2026-02-21 13:46:30 -08:00
Jamie PineandGitHub	68558243d9	Merge pull request #126 from lemassykoi/main Create requirements.txt	2026-02-21 13:46:07 -08:00
Jamie PineandGitHub	8d5ad926f9	Merge pull request #128 from mrigankad/fix/voicebox-bugs fix: resolve multiple issues (#96, #119, #111, #108, #121, #125, #127)	2026-02-21 13:45:19 -08:00
Jamie PineandGitHub	334f037dce	Merge pull request #146 from xPolar/landing/spacebot-banner Add Spacebot banner to landing page	2026-02-21 13:41:31 -08:00
xPolar	f6522eea80	Add Spacebot banner to landing page Adds a persistent top-of-page banner linking to spacebot.sh, another project by the creator of Voicebox. Uses existing design tokens for a consistent look.	2026-02-21 13:37:44 -08:00
lemassykoiandAmp	7615a08f81	ci: add Windows-only build workflow without signing Amp-Thread-ID: https://ampcode.com/threads/T-019c7c3e-072f-7109-86a7-072a6b309891 Co-authored-by: Amp <[email protected]>	2026-02-20 23:06:45 +01:00
lemassykoiandAmp	31ea3c68a5	fix: remove silent browser fallback that bypasses save dialog path Amp-Thread-ID: https://ampcode.com/threads/T-019c7c3e-072f-7109-86a7-072a6b309891 Co-authored-by: Amp <[email protected]>	2026-02-20 19:27:41 +01:00
Mriganka	54d72ddfd0	fix: resolve multiple issues (#96 , #119 , #111 , #108 , #121 , #125 , #127 )	2026-02-20 23:23:38 +05:30
Clément PAPPALARDOandGitHub	d4794f78e1	Create requirements.txt	2026-02-20 17:14:14 +01:00
white1107	aa7c9a9a8d	fix(web): add @tailwindcss/vite plugin to web config The web version was missing the Tailwind CSS Vite plugin, causing CSS to not load at all. This adds the same plugin configuration that exists in the tauri version. Fixes #121	2026-02-20 20:11:27 +09:00
AbrahamandClaude Opus 4.6	ca6ed0998a	Fix model size selection ignored when generating speech The /generate endpoint created the voice prompt before loading the user's requested model size. Since create_voice_prompt() internally calls load_model_async(None), it fell back to the hardcoded default of "1.7B", causing the 1.7B model to be downloaded even when the user explicitly selected 0.6B. This reorders the operations so the requested model is loaded first, ensuring create_voice_prompt() and generate() use the correct model. Co-Authored-By: Claude Opus 4.6 <[email protected]>	2026-02-18 09:41:44 -08:00
Eva	829d4d6d5b	fix(mlx): bundle native libs and broaden error handling for Apple Silicon The distributed macOS aarch64 binary shipped without MLX acceleration despite the model and backend code supporting it. Two root causes: 1. OSError not caught in platform_detect.py PyInstaller bundles isolate the filesystem, so when MLX tries to load its Metal shader libraries (.metallib) it raises OSError, not ImportError. platform_detect.get_backend_type() only caught ImportError, causing a silent fallback to PyTorch even on Apple Silicon hardware. Fix: broaden the except clause to (ImportError, OSError, RuntimeError) and import mlx.core instead of mlx (forces native lib loading eagerly). 2. collect_data_files used instead of collect_all for MLX build_binary.py and voicebox-server.spec used --collect-data / collect_data_files for mlx and mlx_audio. This copies Python source and pure-Python data, but NOT native shared libraries (.dylib, .metallib). Fix: switch to --collect-all / collect_all which captures binaries too, then pass them to Analysis(binaries=...) in the spec. Result: macOS Apple Silicon users now get MLX inference (~4-5x faster than PyTorch CPU), matching the performance documented in the README.	2026-02-18 16:51:48 +01:00
YuriiandCursor	0be7975db5	fix: handle non-ASCII filenames in Content-Disposition headers The export endpoints (export-audio, export generation, export profile, export story) crash with `'latin-1' codec can't encode characters` when the generated text or profile/story name contains non-ASCII characters (e.g. Cyrillic, Chinese, Arabic). Root cause: Python's `str.isalnum()` passes Unicode letters through to the filename, but HTTP headers are encoded as latin-1 by the ASGI server, which cannot represent characters outside the 0-255 range. Fix: introduce `_safe_content_disposition()` helper that builds a standards-compliant header with an ASCII-only `filename` fallback and a RFC 5987 `filename*=UTF-8''...` parameter for Unicode-capable clients. Fixes #68 Co-authored-by: Cursor <[email protected]>	2026-02-17 12:58:42 +04:00
tomasmach	40e4af828a	fix: guard getUserMedia call against undefined mediaDevices in non-secure contexts	2026-02-17 09:28:23 +01:00
Manuel Lorenzo	0e57826ea5	fix: await for confirmation before deleting voices and channels	2026-02-17 00:29:52 +01:00
Spacedrive Mac Mini 2	eb2cd861b1	chore: update Cargo.lock version to 0.1.12	2026-02-10 06:59:41 -08:00
Jamie PineandGitHub	701cc647a7	Merge pull request #57 from selop/chore/readme chore: updates repo URL in README	2026-02-06 05:08:24 -08:00
Sergej Lopatkin	be6ccaf044	chore: updates repo URL in README Updates the repository URL in the README to point to the correct fork. Adds a prerequisite for XCode on macOS for development.	2026-02-06 13:22:59 +01:00
Jamie PineandGitHub	1040625a88	Merge pull request #44 from selop/feature/delivery-instructions Enhances floating generate box UX	2026-02-02 17:58:19 -08:00
Sergej Lopatkin	6f4503b521	Enhances floating generate box UX - Adds tooltips on hover for buttons of the generate box - Replaces the message square icon with a sliders icon for the instruction mode toggle. - Adds a tooltip to the instruction mode toggle button. - Updates the placeholder text for the input field.	2026-02-02 22:19:54 +01:00
Sergej LopatkinandGitHub	f5b6edc2e7	Merge pull request #1 from jamiepine/main update fork	2026-02-02 22:19:30 +01:00
Jamie PineandGitHub	8197f0724c	Merge pull request #40 from Spyabo/fix/audio-export-path-resolution Fix: audio export path resolution	2026-02-02 06:54:39 -08:00
Reese Wright	d40f7d2676	refactor: improve path resolution readability	2026-02-02 14:54:05 +00:00
Reese Wright	99fbcca7f4	update CHANGELOG for audio export fix	2026-02-02 14:34:39 +00:00
Reese Wright	04f9880c9a	fix audio export path resolution	2026-02-02 14:26:34 +00:00
Jamie Pine	b9c858295d	Update Voicebox description as an alternative to ElevenLabs, rather than Ollama	2026-02-01 00:45:47 -08:00
Jamie Pine	610f64c762	fix linux compile	2026-01-31 07:44:28 -08:00
Jamie Pine	220333b3bb	corrections	2026-01-31 02:15:45 -08:00
Jamie Pine	e194e95512	corrections	2026-01-31 02:14:37 -08:00
Jamie Pine	e796412c2c	corrections	2026-01-31 02:13:42 -08:00
Jamie Pine	cb541521d2	Update TTS Provider Architecture status to v0.1.13	2026-01-31 02:11:41 -08:00
Jamie Pine	2bc243f93e	Add TTS Provider Architecture plan Solves GitHub 2GB limit + frequent update UX issues by splitting app into: - Main app (~150MB): UI + backend logic + Whisper - TTS Providers (plugins): Separate downloadable binaries - pytorch-cpu (~300MB) - pytorch-cuda (~2.4GB) - mlx (~800MB, macOS) - remote (connect to external server) - openai (API wrapper) Benefits: - Main app under GitHub 2GB limit - Updates don't require re-downloading providers - User choice of compute backend - External provider support for teams/cloud - Future-proof extensibility	2026-01-31 02:09:42 -08:00
Jamie Pine	0209008d73	disable cuda for 0.1.12	2026-01-31 01:46:14 -08:00