feat: model management improvements and folder migration

- Add model folder migration with byte-level progress tracking (backend + UI)
- Custom models directory support via VOICEBOX_MODELS_DIR env var passed to sidecar
- Hardcoded model descriptions displayed in model detail cards
- Open model folder button in storage location row
- Remove 'not downloaded' badge from model cards
- Fix server settings scroll offset for audio player
- Fix shell open permission to allow file paths
- Add normalize toggle to generation settings

.dockerignore

@@ -0,0 +1,46 @@
+# Version control
+.git
+.github
+.gitignore
+
+# Desktop-only (not needed in web container)
+tauri/
+landing/
+docs/
+mlx-test/
+scripts/
+
+# Dependencies & build artifacts (rebuilt in Docker)
+node_modules/
+__pycache__/
+*.pyc
+*.pyo
+*.egg-info/
+dist/
+build/
+*.spec
+
+# Data (will be bind-mounted)
+data/
+backend/data/
+
+# IDE & OS
+.vscode/
+.idea/
+*.swp
+*.swo
+.DS_Store
+Thumbs.db
+
+# Config files not needed in container
+biome.json
+.biomeignore
+.bumpversion.cfg
+.npmrc
+Makefile
+CHANGELOG.md
+CONTRIBUTING.md
+SECURITY.md
+LICENSE
+README.md
+backend/README.md

.github/workflows/release.yml

@@ -22,10 +22,10 @@ jobs:
             args: "--target x86_64-apple-darwin"
             python-version: "3.12"
             backend: "pytorch"
-          # - platform: 'ubuntu-22.04'
-          #   args: ''
-          #   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"

CONTRIBUTING.md

@@ -27,6 +27,7 @@ Thank you for your interest in contributing to Voicebox! This document provides
   ```bash
   rustc --version  # Check if installed
   ```
+- **[Tauri Prerequisites](https://v2.tauri.app/start/prerequisites)** - Tauri-specific system dependencies (varies by OS).
 
 - **Git** - Version control
 

Dockerfile

@@ -0,0 +1,79 @@
+# ============================================================
+# Voicebox — Local TTS Server with Web UI (CPU)
+# 3-stage build: Frontend → Python deps → Runtime
+# ============================================================
+
+# === Stage 1: Build frontend ===
+FROM oven/bun:1 AS frontend
+
+WORKDIR /build
+
+# Copy workspace config and frontend source
+COPY package.json bun.lock ./
+COPY app/ ./app/
+COPY web/ ./web/
+
+# Strip workspaces not needed for web build, and fix trailing comma
+RUN sed -i '/"tauri"/d; /"landing"/d' package.json && \
+    sed -i -z 's/,\n  ]/\n  ]/' package.json
+RUN bun install --no-save
+# Build frontend (skip tsc — upstream has pre-existing type errors)
+RUN cd web && bunx --bun vite build
+
+
+# === Stage 2: Build Python dependencies ===
+FROM python:3.11-slim AS backend-builder
+
+WORKDIR /build
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    git \
+    build-essential \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY backend/requirements.txt .
+RUN pip install --no-cache-dir --prefix=/install -r requirements.txt
+RUN pip install --no-cache-dir --prefix=/install \
+    git+https://github.com/QwenLM/Qwen3-TTS.git
+
+
+# === Stage 3: Runtime ===
+FROM python:3.11-slim
+
+# Create non-root user for security
+RUN groupadd -r voicebox && \
+    useradd -r -g voicebox -m -s /bin/bash voicebox
+
+WORKDIR /app
+
+# Install only runtime system dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ffmpeg \
+    curl \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy installed Python packages from builder stage
+COPY --from=backend-builder /install /usr/local
+
+# Copy backend application code
+COPY --chown=voicebox:voicebox backend/ /app/backend/
+
+# Copy built frontend from frontend stage
+COPY --from=frontend --chown=voicebox:voicebox /build/web/dist /app/frontend/
+
+# Create data directories owned by non-root user
+RUN mkdir -p /app/data/generations /app/data/profiles /app/data/cache \
+    && chown -R voicebox:voicebox /app/data
+
+# Switch to non-root user
+USER voicebox
+
+# Expose the API port
+EXPOSE 17493
+
+# Health check — auto-restart if the server hangs
+HEALTHCHECK --interval=30s --timeout=10s --retries=3 --start-period=60s \
+    CMD curl -f http://localhost:17493/health || exit 1
+
+# Start the FastAPI server
+CMD ["uvicorn", "backend.main:app", "--host", "0.0.0.0", "--port", "17493"]

PATCH_NOTES.md

@@ -0,0 +1,58 @@
+# Voicebox Offline Mode Fix
+
+## Problem
+Voicebox crashes when generating speech if HuggingFace is unreachable, even when models are fully cached locally.
+
+**Root Cause:**
+- Voicebox downloads `mlx-community/Qwen3-TTS-12Hz-1.7B-Base-bf16` (MLX optimized version)
+- But `mlx_audio.tts.load()` tries to fetch `config.json` from original repo `Qwen/Qwen3-TTS-12Hz-1.7B-Base`
+- This network request fails → server crashes with `RemoteDisconnected`
+
+**Related Issues:**
+- Issue #150: "Internet connection required, even though models are downloaded?"
+- Issue #151: "API Stability Issues: Model Loading Hangs and Server Crashes"
+
+## Solution
+Two-part fix:
+
+### 1. Monkey-patch huggingface_hub (`backend/utils/hf_offline_patch.py`)
+- Intercepts cache lookup functions
+- Forces offline mode early (before mlx_audio imports)
+- Adds debug logging for cache hits/misses
+
+### 2. Symlink original repo to MLX version (`ensure_original_qwen_config_cached()`)
+- When original `Qwen/Qwen3-TTS-12Hz-1.7B-Base` cache doesn't exist
+- But MLX `mlx-community/Qwen3-TTS-12Hz-1.7B-Base-bf16` does exist
+- Creates a symlink so cache lookups succeed
+
+## Files Changed
+- `backend/backends/mlx_backend.py` - Added patch imports at top
+- `backend/utils/hf_offline_patch.py` - New patch module
+
+## Testing
+To test this fix:
+1. Build Voicebox from source: `make build`
+2. Disconnect from internet
+3. Try generating speech
+4. Should work without network requests
+
+## Build Instructions
+
+```bash
+# Install dependencies
+pip install -r requirements.txt
+
+# Build the app
+make build
+
+# Or build just the server
+make build-server
+```
+
+## Notes
+- The patch is applied automatically when `mlx_backend.py` is imported
+- Set `VOICEBOX_OFFLINE_PATCH=0` to disable the patch
+- The symlink approach works because the config.json is compatible between versions
+
+---
+*Patch contributed by community*

README.md

@@ -98,12 +98,12 @@ Powered by Alibaba's **Qwen3-TTS** — a breakthrough model that achieves near-p
 - **Instant cloning** — Upload a sample, get a voice profile
 - **High fidelity** — Natural prosody, emotion, and cadence
 - **Multi-language** — English, Chinese, and more coming
-- **Lightning fast on Mac** — MLX backend leverages Apple Silicon's Neural Engine for super fast generation
+- **Lightning fast on Mac** — MLX backend leverages Apple Silicon's Neural Engine for super-fast generation
 
 ### Voice Profile Management
 
 - **Create profiles** from audio files or record directly in-app
-- **Import/Export** profiles to share or backup
+- **Import/Export** profiles to share or back up
 - **Multi-sample support** — combine multiple samples for higher quality cloning
 - **Organize** with descriptions and language tags
 
@@ -242,7 +242,7 @@ Install [just](https://github.com/casey/just): `brew install just` or `cargo ins
 
 Also available via Makefile: `make setup && make dev` (run `make help` for all commands).
 
-**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org), [XCode on macOS](https://developer.apple.com/xcode/).
+**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org), [XCode on macOS](https://developer.apple.com/xcode/), [Tauri Prerequisites](https://v2.tauri.app/start/prerequisites/).
 
 **Performance:** 
 - **Apple Silicon (M1/M2/M3)**: Uses MLX backend with native Metal acceleration for 4-5x faster inference

app/src/App.tsx

@@ -93,10 +93,12 @@ function App() {
     }
 
     serverStartingRef.current = true;
-    console.log('Production mode: Starting bundled server...');
+    const isRemote = useServerStore.getState().mode === 'remote';
+    const customModelsDir = useServerStore.getState().customModelsDir;
+    console.log(`Production mode: Starting bundled server... (remote: ${isRemote})`);
 
     platform.lifecycle
-      .startServer(false)
+      .startServer(isRemote, customModelsDir)
       .then((serverUrl) => {
         console.log('Server is ready at:', serverUrl);
         // Update the server URL in the store with the dynamically assigned port

app/src/components/AudioTab/AudioTab.tsx

@@ -23,8 +23,8 @@ import {
 import { apiClient } from '@/lib/api/client';
 import { BOTTOM_SAFE_AREA_PADDING } from '@/lib/constants/ui';
 import { cn } from '@/lib/utils/cn';
-import { usePlayerStore } from '@/stores/playerStore';
 import { usePlatform } from '@/platform/PlatformContext';
+import { usePlayerStore } from '@/stores/playerStore';
 
 interface AudioDevice {
   id: string;
@@ -129,7 +129,7 @@ export function AudioTab() {
     if (await confirm('Delete this channel?')) {
       deleteChannel.mutate(channelId);
     }
-  }
+  };
 
   const allChannels = channels || [];
   const allDevices = devices || [];
@@ -168,7 +168,7 @@ export function AudioTab() {
               </Button>
             </div>
           ) : (
-            <div className="space-y-3 p-2">
+            <div className="space-y-3">
               {allChannels.map((channel) => {
                 const isSelected = selectedChannelId === channel.id;
                 return (
@@ -343,7 +343,9 @@ export function AudioTab() {
             <div className="flex flex-col items-center justify-center py-12 border-2 border-dashed border-muted rounded-md">
               <CheckCircle2 className="h-12 w-12 text-muted-foreground mb-4" />
               <p className="text-muted-foreground text-center">
-                {platform.metadata.isTauri ? 'No audio devices found' : 'Audio device selection requires Tauri'}
+                {platform.metadata.isTauri
+                  ? 'No audio devices found'
+                  : 'Audio device selection requires Tauri'}
               </p>
             </div>
           )}

app/src/components/Generation/FloatingGenerateBox.tsx

@@ -13,13 +13,14 @@ import {
 } from '@/components/ui/select';
 import { Textarea } from '@/components/ui/textarea';
 import { useToast } from '@/components/ui/use-toast';
-import { LANGUAGE_OPTIONS } from '@/lib/constants/languages';
+import { getLanguageOptionsForEngine, type LanguageCode } from '@/lib/constants/languages';
 import { useGenerationForm } from '@/lib/hooks/useGenerationForm';
 import { useProfile, useProfiles } from '@/lib/hooks/useProfiles';
 import { useAddStoryItem, useStory } from '@/lib/hooks/useStories';
 import { cn } from '@/lib/utils/cn';
 import { useStoryStore } from '@/stores/storyStore';
 import { useUIStore } from '@/stores/uiStore';
+import { ParalinguisticInput } from './ParalinguisticInput';
 
 interface FloatingGenerateBoxProps {
   isPlayerOpen?: boolean;
@@ -112,6 +113,13 @@ export function FloatingGenerateBox({
     }
   }, [selectedProfileId, profiles, setSelectedProfileId]);
 
+  // Sync generation form language with selected profile's language
+  useEffect(() => {
+    if (selectedProfile?.language) {
+      form.setValue('language', selectedProfile.language as LanguageCode);
+    }
+  }, [selectedProfile, form]);
+
   // Auto-resize textarea based on content (only when expanded)
   useEffect(() => {
     if (!isExpanded) {
@@ -212,34 +220,57 @@ export function FloatingGenerateBox({
                             transition={{ duration: 0.15, ease: 'easeOut' }}
                             style={{ overflow: 'hidden' }}
                           >
-                            <Textarea
-                              {...field}
-                              ref={(node: HTMLTextAreaElement | null) => {
-                                // Store ref for auto-resize (only for active field)
-                                if (!isInstructMode) {
-                                  textareaRef.current = node;
+                            {form.watch('engine') === 'chatterbox_turbo' ? (
+                              <ParalinguisticInput
+                                value={field.value}
+                                onChange={field.onChange}
+                                placeholder={
+                                  isStoriesRoute && currentStory
+                                    ? `Generate speech for "${currentStory.name}"... (type / for effects)`
+                                    : selectedProfile
+                                      ? `Type / for effects like [laugh], [sigh]...`
+                                      : 'Select a voice profile above...'
                                 }
-                                // Forward ref to react-hook-form
-                                if (typeof field.ref === 'function') {
-                                  field.ref(node);
+                                className="px-3 py-2 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 w-full"
+                                style={{
+                                  minHeight: isExpanded ? '100px' : '32px',
+                                  maxHeight: '300px',
+                                  overflowY: 'auto',
+                                }}
+                                disabled={!selectedProfileId}
+                                onClick={() => setIsExpanded(true)}
+                                onFocus={() => setIsExpanded(true)}
+                              />
+                            ) : (
+                              <Textarea
+                                {...field}
+                                ref={(node: HTMLTextAreaElement | null) => {
+                                  // Store ref for auto-resize (only for active field)
+                                  if (!isInstructMode) {
+                                    textareaRef.current = node;
+                                  }
+                                  // Forward ref to react-hook-form
+                                  if (typeof field.ref === 'function') {
+                                    field.ref(node);
+                                  }
+                                }}
+                                placeholder={
+                                  isStoriesRoute && currentStory
+                                    ? `Generate speech for "${currentStory.name}"...`
+                                    : selectedProfile
+                                      ? `Generate speech using ${selectedProfile.name}...`
+                                      : 'Select a voice profile above...'
                                 }
-                              }}
-                              placeholder={
-                                isStoriesRoute && currentStory
-                                  ? `Generate speech for "${currentStory.name}"...`
-                                  : selectedProfile
-                                    ? `Generate speech using ${selectedProfile.name}...`
-                                    : 'Select a voice profile above...'
-                              }
-                              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',
-                                maxHeight: '300px',
-                              }}
-                              disabled={!selectedProfileId}
-                              onClick={() => setIsExpanded(true)}
-                              onFocus={() => setIsExpanded(true)}
-                            />
+                                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',
+                                  maxHeight: '300px',
+                                }}
+                                disabled={!selectedProfileId}
+                                onClick={() => setIsExpanded(true)}
+                                onFocus={() => setIsExpanded(true)}
+                              />
+                            )}
                           </motion.div>
                         </FormControl>
                         <FormMessage className="text-xs" />
@@ -344,9 +375,7 @@ export function FloatingGenerateBox({
                               : 'bg-card border border-border hover:bg-background/50',
                           )}
                           aria-label={
-                            isInstructMode
-                              ? 'Fine tune instructions, on'
-                              : 'Fine tune instructions'
+                            isInstructMode ? 'Fine tune instructions, on' : 'Fine tune instructions'
                           }
                         >
                           <SlidersHorizontal className="h-4 w-4" />
@@ -393,25 +422,30 @@ export function FloatingGenerateBox({
                   <FormField
                     control={form.control}
                     name="language"
-                    render={({ field }) => (
-                      <FormItem className="flex-1 space-y-0">
-                        <Select onValueChange={field.onChange} defaultValue={field.value}>
-                          <FormControl>
-                            <SelectTrigger className="h-8 text-xs bg-card border-border rounded-full hover:bg-background/50 transition-all">
-                              <SelectValue />
-                            </SelectTrigger>
-                          </FormControl>
-                          <SelectContent>
-                            {LANGUAGE_OPTIONS.map((lang) => (
-                              <SelectItem key={lang.value} value={lang.value} className="text-xs">
-                                {lang.label}
-                              </SelectItem>
-                            ))}
-                          </SelectContent>
-                        </Select>
-                        <FormMessage className="text-xs" />
-                      </FormItem>
-                    )}
+                    render={({ field }) => {
+                      const engineLangs = getLanguageOptionsForEngine(
+                        form.watch('engine') || 'qwen',
+                      );
+                      return (
+                        <FormItem className="flex-1 space-y-0">
+                          <Select onValueChange={field.onChange} value={field.value}>
+                            <FormControl>
+                              <SelectTrigger className="h-8 text-xs bg-card border-border rounded-full hover:bg-background/50 transition-all">
+                                <SelectValue />
+                              </SelectTrigger>
+                            </FormControl>
+                            <SelectContent>
+                              {engineLangs.map((lang) => (
+                                <SelectItem key={lang.value} value={lang.value} className="text-xs">
+                                  {lang.label}
+                                </SelectItem>
+                              ))}
+                            </SelectContent>
+                          </Select>
+                          <FormMessage className="text-xs" />
+                        </FormItem>
+                      );
+                    }}
                   />
 
                   <FormItem className="flex-1 space-y-0">
@@ -421,13 +455,19 @@ export function FloatingGenerateBox({
                           ? 'luxtts'
                           : form.watch('engine') === 'chatterbox'
                             ? 'chatterbox'
-                            : `qwen:${form.watch('modelSize') || '1.7B'}`
+                            : form.watch('engine') === 'chatterbox_turbo'
+                              ? 'chatterbox_turbo'
+                              : `qwen:${form.watch('modelSize') || '1.7B'}`
                       }
                       onValueChange={(value) => {
                         if (value === 'luxtts') {
                           form.setValue('engine', 'luxtts');
+                          form.setValue('language', 'en');
                         } else if (value === 'chatterbox') {
                           form.setValue('engine', 'chatterbox');
+                        } else if (value === 'chatterbox_turbo') {
+                          form.setValue('engine', 'chatterbox_turbo');
+                          form.setValue('language', 'en');
                         } else {
                           const [, modelSize] = value.split(':');
                           form.setValue('engine', 'qwen');
@@ -453,6 +493,12 @@ export function FloatingGenerateBox({
                         <SelectItem value="chatterbox" className="text-xs text-muted-foreground">
                           Chatterbox
                         </SelectItem>
+                        <SelectItem
+                          value="chatterbox_turbo"
+                          className="text-xs text-muted-foreground"
+                        >
+                          Chatterbox Turbo
+                        </SelectItem>
                       </SelectContent>
                     </Select>
                   </FormItem>

app/src/components/Generation/GenerationForm.tsx

@@ -19,10 +19,11 @@ import {
   SelectValue,
 } from '@/components/ui/select';
 import { Textarea } from '@/components/ui/textarea';
-import { LANGUAGE_OPTIONS } from '@/lib/constants/languages';
+import { getLanguageOptionsForEngine } from '@/lib/constants/languages';
 import { useGenerationForm } from '@/lib/hooks/useGenerationForm';
 import { useProfile } from '@/lib/hooks/useProfiles';
 import { useUIStore } from '@/stores/uiStore';
+import { ParalinguisticInput } from './ParalinguisticInput';
 
 export function GenerationForm() {
   const selectedProfileId = useUIStore((state) => state.selectedProfileId);
@@ -64,13 +65,26 @@ export function GenerationForm() {
                 <FormItem>
                   <FormLabel>Text to Speak</FormLabel>
                   <FormControl>
-                    <Textarea
-                      placeholder="Enter the text you want to generate..."
-                      className="min-h-[150px]"
-                      {...field}
-                    />
+                    {form.watch('engine') === 'chatterbox_turbo' ? (
+                      <ParalinguisticInput
+                        value={field.value}
+                        onChange={field.onChange}
+                        placeholder="Enter text... type / for effects like [laugh], [sigh]"
+                        className="min-h-[150px] rounded-md border border-input bg-background px-3 py-2"
+                      />
+                    ) : (
+                      <Textarea
+                        placeholder="Enter the text you want to generate..."
+                        className="min-h-[150px]"
+                        {...field}
+                      />
+                    )}
                   </FormControl>
-                  <FormDescription>Max 5000 characters</FormDescription>
+                  <FormDescription>
+                    {form.watch('engine') === 'chatterbox_turbo'
+                      ? 'Max 5000 characters. Type / to insert sound effects.'
+                      : 'Max 5000 characters'}
+                  </FormDescription>
                   <FormMessage />
                 </FormItem>
               )}
@@ -109,13 +123,19 @@ export function GenerationForm() {
                       ? 'luxtts'
                       : form.watch('engine') === 'chatterbox'
                         ? 'chatterbox'
-                        : `qwen:${form.watch('modelSize') || '1.7B'}`
+                        : form.watch('engine') === 'chatterbox_turbo'
+                          ? 'chatterbox_turbo'
+                          : `qwen:${form.watch('modelSize') || '1.7B'}`
                   }
                   onValueChange={(value) => {
                     if (value === 'luxtts') {
                       form.setValue('engine', 'luxtts');
+                      form.setValue('language', 'en');
                     } else if (value === 'chatterbox') {
                       form.setValue('engine', 'chatterbox');
+                    } else if (value === 'chatterbox_turbo') {
+                      form.setValue('engine', 'chatterbox_turbo');
+                      form.setValue('language', 'en');
                     } else {
                       const [, modelSize] = value.split(':');
                       form.setValue('engine', 'qwen');
@@ -133,40 +153,46 @@ export function GenerationForm() {
                     <SelectItem value="qwen:0.6B">Qwen3-TTS 0.6B</SelectItem>
                     <SelectItem value="luxtts">LuxTTS</SelectItem>
                     <SelectItem value="chatterbox">Chatterbox</SelectItem>
+                    <SelectItem value="chatterbox_turbo">Chatterbox Turbo</SelectItem>
                   </SelectContent>
                 </Select>
                 <FormDescription>
                   {form.watch('engine') === 'luxtts'
                     ? 'Fast, English-focused'
                     : form.watch('engine') === 'chatterbox'
-                      ? 'Multilingual, incl. Hebrew'
-                      : 'Multi-language, two sizes'}
+                      ? '23 languages, incl. Hebrew'
+                      : form.watch('engine') === 'chatterbox_turbo'
+                        ? 'English, [laugh] [cough] tags'
+                        : 'Multi-language, two sizes'}
                 </FormDescription>
               </FormItem>
 
               <FormField
                 control={form.control}
                 name="language"
-                render={({ field }) => (
-                  <FormItem>
-                    <FormLabel>Language</FormLabel>
-                    <Select onValueChange={field.onChange} defaultValue={field.value}>
-                      <FormControl>
-                        <SelectTrigger>
-                          <SelectValue />
-                        </SelectTrigger>
-                      </FormControl>
-                      <SelectContent>
-                        {LANGUAGE_OPTIONS.map((lang) => (
-                          <SelectItem key={lang.value} value={lang.value}>
-                            {lang.label}
-                          </SelectItem>
-                        ))}
-                      </SelectContent>
-                    </Select>
-                    <FormMessage />
-                  </FormItem>
-                )}
+                render={({ field }) => {
+                  const engineLangs = getLanguageOptionsForEngine(form.watch('engine') || 'qwen');
+                  return (
+                    <FormItem>
+                      <FormLabel>Language</FormLabel>
+                      <Select onValueChange={field.onChange} value={field.value}>
+                        <FormControl>
+                          <SelectTrigger>
+                            <SelectValue />
+                          </SelectTrigger>
+                        </FormControl>
+                        <SelectContent>
+                          {engineLangs.map((lang) => (
+                            <SelectItem key={lang.value} value={lang.value}>
+                              {lang.label}
+                            </SelectItem>
+                          ))}
+                        </SelectContent>
+                      </Select>
+                      <FormMessage />
+                    </FormItem>
+                  );
+                }}
               />
 
               <FormField

app/src/components/Generation/ParalinguisticInput.tsx

@@ -0,0 +1,422 @@
+/**
+ * ParalinguisticInput — a contentEditable rich text input that renders
+ * Chatterbox Turbo paralinguistic tags (e.g. [laugh]) as inline badges.
+ *
+ * Trigger: typing "/" opens an autocomplete dropdown.
+ * Paste:   pasting text with [tag] patterns auto-converts to badges.
+ * Output:  serializes badges back to plain [tag] text for the API.
+ */
+
+import { AnimatePresence, motion } from 'framer-motion';
+import { forwardRef, useCallback, useEffect, useImperativeHandle, useRef, useState } from 'react';
+import { createPortal } from 'react-dom';
+import { cn } from '@/lib/utils/cn';
+
+// ── Tag definitions ─────────────────────────────────────────────────
+const PARALINGUISTIC_TAGS = [
+  { tag: '[laugh]', label: 'laugh', emoji: '\u{1F602}' },
+  { tag: '[chuckle]', label: 'chuckle', emoji: '\u{1F60F}' },
+  { tag: '[gasp]', label: 'gasp', emoji: '\u{1F62E}' },
+  { tag: '[cough]', label: 'cough', emoji: '\u{1F637}' },
+  { tag: '[sigh]', label: 'sigh', emoji: '\u{1F614}' },
+  { tag: '[groan]', label: 'groan', emoji: '\u{1F629}' },
+  { tag: '[sniff]', label: 'sniff', emoji: '\u{1F443}' },
+  { tag: '[shush]', label: 'shush', emoji: '\u{1F92B}' },
+  { tag: '[clear throat]', label: 'clear throat', emoji: '\u{1F64A}' },
+] as const;
+
+const TAG_REGEX = /\[(laugh|chuckle|gasp|cough|sigh|groan|sniff|shush|clear throat)\]/gi;
+
+// Data attribute used to identify badge spans in the DOM
+const BADGE_ATTR = 'data-ptag';
+
+// ── Helpers ─────────────────────────────────────────────────────────
+
+/** Build an inline badge <span> for a tag. */
+function makeBadgeHTML(tag: string): string {
+  const entry = PARALINGUISTIC_TAGS.find((t) => t.tag.toLowerCase() === tag.toLowerCase());
+  const label = entry?.label ?? tag.replace(/[[\]]/g, '');
+  const emoji = entry?.emoji ?? '';
+  // Non-editable inline badge. Zero-width spaces around it let the
+  // caret sit on either side so the user can type before/after.
+  return `\u200B<span ${BADGE_ATTR}="${tag}" contenteditable="false" class="ptag-badge">${emoji ? `${emoji}\u00A0` : ''}${label}</span>\u200B`;
+}
+
+/** Convert plain text with [tag] patterns into HTML with badge spans. */
+function textToHTML(text: string): string {
+  // Escape HTML entities first
+  const escaped = text.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;');
+  // Replace tag patterns with badge HTML
+  return escaped.replace(TAG_REGEX, (match) => makeBadgeHTML(match));
+}
+
+/** Serialize the contentEditable innerHTML back to plain text with [tag] syntax. */
+function htmlToText(container: HTMLElement): string {
+  let result = '';
+  for (const node of container.childNodes) {
+    if (node.nodeType === Node.TEXT_NODE) {
+      // Strip zero-width spaces we added around badges
+      result += (node.textContent ?? '').replace(/\u200B/g, '');
+    } else if (node.nodeType === Node.ELEMENT_NODE) {
+      const el = node as HTMLElement;
+      if (el.hasAttribute(BADGE_ATTR)) {
+        result += el.getAttribute(BADGE_ATTR) ?? '';
+      } else if (el.tagName === 'BR') {
+        result += '\n';
+      } else {
+        // Recurse for nested elements (e.g. spans from paste)
+        result += htmlToText(el);
+      }
+    }
+  }
+  return result;
+}
+
+/** Get the text content from the current caret position back to the last
+ *  whitespace or start of container, to detect the "/" trigger. */
+function getWordBeforeCaret(_container: HTMLElement): { word: string; range: Range | null } {
+  const sel = window.getSelection();
+  if (!sel || sel.rangeCount === 0) return { word: '', range: null };
+  const range = sel.getRangeAt(0).cloneRange();
+  range.collapse(true);
+
+  // Walk backwards from caret through the text node
+  const textNode = range.startContainer;
+  if (textNode.nodeType !== Node.TEXT_NODE) return { word: '', range: null };
+  const text = textNode.textContent ?? '';
+  const offset = range.startOffset;
+
+  let start = offset;
+  while (
+    start > 0 &&
+    text[start - 1] !== ' ' &&
+    text[start - 1] !== '\n' &&
+    text[start - 1] !== '\u00A0'
+  ) {
+    start--;
+  }
+
+  const word = text.slice(start, offset);
+  const wordRange = document.createRange();
+  wordRange.setStart(textNode, start);
+  wordRange.setEnd(textNode, offset);
+
+  return { word, range: wordRange };
+}
+
+// ── Component ───────────────────────────────────────────────────────
+
+export interface ParalinguisticInputProps {
+  value?: string;
+  onChange?: (value: string) => void;
+  placeholder?: string;
+  disabled?: boolean;
+  className?: string;
+  style?: React.CSSProperties;
+  onClick?: () => void;
+  onFocus?: () => void;
+}
+
+export interface ParalinguisticInputRef {
+  focus: () => void;
+  element: HTMLDivElement | null;
+}
+
+export const ParalinguisticInput = forwardRef<ParalinguisticInputRef, ParalinguisticInputProps>(
+  function ParalinguisticInput(
+    { value, onChange, placeholder, disabled, className, style, onClick, onFocus },
+    ref,
+  ) {
+    const editorRef = useRef<HTMLDivElement>(null);
+    const [showMenu, setShowMenu] = useState(false);
+    const [menuFilter, setMenuFilter] = useState('');
+    const [menuIndex, setMenuIndex] = useState(0);
+    const [menuPosition, setMenuPosition] = useState<{ bottom: number; left: number }>({
+      bottom: 0,
+      left: 0,
+    });
+    const triggerRangeRef = useRef<Range | null>(null);
+    const lastSerializedRef = useRef<string>('');
+    const isComposingRef = useRef(false);
+
+    useImperativeHandle(ref, () => ({
+      focus: () => editorRef.current?.focus(),
+      element: editorRef.current,
+    }));
+
+    // Filtered tag list for the autocomplete menu
+    const filteredTags = PARALINGUISTIC_TAGS.filter((t) =>
+      t.label.toLowerCase().includes(menuFilter.toLowerCase()),
+    );
+
+    // ── Sync external value → editor ──────────────────────────────
+    useEffect(() => {
+      const el = editorRef.current;
+      if (!el) return;
+      // Only update DOM if the external value differs from what we last emitted
+      if (value !== undefined && value !== lastSerializedRef.current) {
+        lastSerializedRef.current = value;
+        el.innerHTML = value ? textToHTML(value) : '';
+      }
+    }, [value]);
+
+    // ── Emit plain-text value on input ────────────────────────────
+    const emitChange = useCallback(() => {
+      const el = editorRef.current;
+      if (!el || !onChange) return;
+      const text = htmlToText(el);
+      lastSerializedRef.current = text;
+      onChange(text);
+    }, [onChange]);
+
+    // ── Insert a tag badge at the caret ───────────────────────────
+    const insertTag = useCallback(
+      (tag: string) => {
+        const el = editorRef.current;
+        if (!el) return;
+
+        // Delete the /filter text
+        const wordRange = triggerRangeRef.current;
+        if (wordRange) {
+          wordRange.deleteContents();
+        }
+
+        // Insert badge HTML
+        const temp = document.createElement('span');
+        temp.innerHTML = makeBadgeHTML(tag);
+        const frag = document.createDocumentFragment();
+        let lastNode: Node | null = null;
+        while (temp.firstChild) {
+          lastNode = frag.appendChild(temp.firstChild);
+        }
+
+        const sel = window.getSelection();
+        if (sel && sel.rangeCount > 0) {
+          const range = sel.getRangeAt(0);
+          range.deleteContents();
+          range.insertNode(frag);
+
+          // Move caret after the badge
+          if (lastNode) {
+            const newRange = document.createRange();
+            newRange.setStartAfter(lastNode);
+            newRange.collapse(true);
+            sel.removeAllRanges();
+            sel.addRange(newRange);
+          }
+        }
+
+        setShowMenu(false);
+        setMenuFilter('');
+        emitChange();
+        el.focus();
+      },
+      [emitChange],
+    );
+
+    // ── Handle keydown for autocomplete navigation ────────────────
+    const handleKeyDown = useCallback(
+      (e: React.KeyboardEvent) => {
+        if (showMenu) {
+          if (filteredTags.length === 0) {
+            if (e.key === 'Escape') {
+              e.preventDefault();
+              setShowMenu(false);
+            }
+            return;
+          }
+          if (e.key === 'ArrowDown') {
+            e.preventDefault();
+            setMenuIndex((i) => (i + 1) % filteredTags.length);
+          } else if (e.key === 'ArrowUp') {
+            e.preventDefault();
+            setMenuIndex((i) => (i - 1 + filteredTags.length) % filteredTags.length);
+          } else if (e.key === 'Enter' || e.key === 'Tab') {
+            e.preventDefault();
+            if (filteredTags[menuIndex]) {
+              insertTag(filteredTags[menuIndex].tag);
+            }
+          } else if (e.key === 'Escape') {
+            e.preventDefault();
+            setShowMenu(false);
+          }
+        } else {
+          // Prevent Enter from creating <div> blocks in contentEditable
+          if (e.key === 'Enter' && !e.shiftKey) {
+            // Let the form handle submit
+          }
+        }
+      },
+      [showMenu, filteredTags, menuIndex, insertTag],
+    );
+
+    // ── Handle input (check for / trigger) ────────────────────────
+    const handleInput = useCallback(() => {
+      if (isComposingRef.current) return;
+      const el = editorRef.current;
+      if (!el) return;
+
+      const { word, range } = getWordBeforeCaret(el);
+
+      if (word.startsWith('/')) {
+        const filter = word.slice(1); // strip the /
+        setMenuFilter(filter);
+        setMenuIndex(0);
+        triggerRangeRef.current = range;
+
+        // Position the menu above the caret using viewport coords (portalled)
+        const sel = window.getSelection();
+        if (sel && sel.rangeCount > 0) {
+          const rect = sel.getRangeAt(0).getBoundingClientRect();
+          setMenuPosition({
+            bottom: window.innerHeight - rect.top + 4,
+            left: rect.left,
+          });
+        }
+
+        setShowMenu(true);
+      } else {
+        setShowMenu(false);
+      }
+
+      emitChange();
+    }, [emitChange]);
+
+    // ── Handle paste — convert [tag] patterns to badges ───────────
+    const handlePaste = useCallback(
+      (e: React.ClipboardEvent) => {
+        e.preventDefault();
+        const text = e.clipboardData.getData('text/plain');
+        if (!text) return;
+
+        const el = editorRef.current;
+        if (!el) return;
+
+        const html = textToHTML(text);
+
+        // Insert at caret
+        const sel = window.getSelection();
+        if (sel && sel.rangeCount > 0) {
+          const range = sel.getRangeAt(0);
+          range.deleteContents();
+          const temp = document.createElement('div');
+          temp.innerHTML = html;
+          const frag = document.createDocumentFragment();
+          let lastNode: Node | null = null;
+          while (temp.firstChild) {
+            lastNode = frag.appendChild(temp.firstChild);
+          }
+          range.insertNode(frag);
+          if (lastNode) {
+            const newRange = document.createRange();
+            newRange.setStartAfter(lastNode);
+            newRange.collapse(true);
+            sel.removeAllRanges();
+            sel.addRange(newRange);
+          }
+        }
+
+        emitChange();
+      },
+      [emitChange],
+    );
+
+    // ── Show placeholder ──────────────────────────────────────────
+    const isEmpty = !value || value.trim() === '';
+
+    return (
+      <div className="relative">
+        {/* Placeholder */}
+        {isEmpty && placeholder && (
+          <div
+            className="pointer-events-none absolute inset-0 text-sm text-muted-foreground/60 px-3 py-2 select-none"
+            aria-hidden
+          >
+            {placeholder}
+          </div>
+        )}
+
+        {/* Editable area */}
+        <div
+          ref={editorRef}
+          contentEditable={!disabled}
+          suppressContentEditableWarning
+          role={disabled ? undefined : 'textbox'}
+          aria-multiline={disabled ? undefined : true}
+          aria-placeholder={placeholder}
+          aria-disabled={disabled}
+          tabIndex={disabled ? -1 : 0}
+          className={cn(
+            'min-h-[32px] text-sm whitespace-pre-wrap break-words outline-none',
+            '[&_.ptag-badge]:inline-flex [&_.ptag-badge]:items-center [&_.ptag-badge]:rounded-full',
+            '[&_.ptag-badge]:bg-accent/20 [&_.ptag-badge]:text-accent [&_.ptag-badge]:border [&_.ptag-badge]:border-accent/30',
+            '[&_.ptag-badge]:px-2 [&_.ptag-badge]:py-0 [&_.ptag-badge]:text-xs [&_.ptag-badge]:font-medium',
+            '[&_.ptag-badge]:mx-0.5 [&_.ptag-badge]:select-none [&_.ptag-badge]:cursor-default',
+            '[&_.ptag-badge]:align-baseline',
+            disabled && 'opacity-50 cursor-not-allowed',
+            className,
+          )}
+          style={style}
+          onInput={!disabled ? handleInput : undefined}
+          onKeyDown={!disabled ? handleKeyDown : undefined}
+          onPaste={!disabled ? handlePaste : undefined}
+          onClick={!disabled ? onClick : undefined}
+          onFocus={!disabled ? onFocus : undefined}
+          onBlur={() => {
+            setShowMenu(false);
+            triggerRangeRef.current = null;
+          }}
+          onCompositionStart={() => {
+            isComposingRef.current = true;
+          }}
+          onCompositionEnd={() => {
+            isComposingRef.current = false;
+            handleInput();
+          }}
+        />
+
+        {/* Autocomplete dropdown — portalled to body, positioned above the caret */}
+        {showMenu &&
+          filteredTags.length > 0 &&
+          createPortal(
+            <AnimatePresence>
+              <motion.div
+                initial={{ opacity: 0, y: 4 }}
+                animate={{ opacity: 1, y: 0 }}
+                exit={{ opacity: 0, y: 4 }}
+                transition={{ duration: 0.12 }}
+                className="fixed z-[9999] min-w-[200px] max-h-[280px] overflow-y-auto rounded-lg border border-border bg-popover shadow-lg"
+                style={{
+                  bottom: menuPosition.bottom,
+                  left: menuPosition.left,
+                }}
+              >
+                {filteredTags.map((t, i) => (
+                  <button
+                    key={t.tag}
+                    type="button"
+                    className={cn(
+                      'flex items-center gap-2 w-full px-3 py-1.5 text-sm text-left transition-colors',
+                      i === menuIndex
+                        ? 'bg-accent/20 text-accent-foreground'
+                        : 'text-popover-foreground hover:bg-muted/50',
+                    )}
+                    onMouseDown={(e) => {
+                      e.preventDefault(); // Keep focus in editor
+                      insertTag(t.tag);
+                    }}
+                    onMouseEnter={() => setMenuIndex(i)}
+                  >
+                    <span className="text-base leading-none">{t.emoji}</span>
+                    <span>{t.label}</span>
+                    <span className="ml-auto text-xs text-muted-foreground font-mono">{t.tag}</span>
+                  </button>
+                ))}
+              </motion.div>
+            </AnimatePresence>,
+            document.body,
+          )}
+      </div>
+    );
+  },
+);

app/src/components/ModelsTab/ModelsTab.tsx

@@ -2,7 +2,7 @@ import { ModelManagement } from '@/components/ServerSettings/ModelManagement';
 
 export function ModelsTab() {
   return (
-    <div className="h-full flex flex-col p-4">
+    <div className="h-full flex flex-col">
       <ModelManagement />
     </div>
   );

app/src/components/ServerSettings/ConnectionForm.tsx

@@ -1,9 +1,12 @@
 import { zodResolver } from '@hookform/resolvers/zod';
+import { Loader2, XCircle } from 'lucide-react';
 import { useEffect } from 'react';
 import { useForm } from 'react-hook-form';
 import * as z from 'zod';
+import { Badge } from '@/components/ui/badge';
 import { Button } from '@/components/ui/button';
 import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card';
+import { Checkbox } from '@/components/ui/checkbox';
 import {
   Form,
   FormControl,
@@ -14,10 +17,10 @@ import {
   FormMessage,
 } from '@/components/ui/form';
 import { Input } from '@/components/ui/input';
-import { Checkbox } from '@/components/ui/checkbox';
 import { useToast } from '@/components/ui/use-toast';
-import { useServerStore } from '@/stores/serverStore';
+import { useServerHealth } from '@/lib/hooks/useServer';
 import { usePlatform } from '@/platform/PlatformContext';
+import { useServerStore } from '@/stores/serverStore';
 
 const connectionSchema = z.object({
   serverUrl: z.string().url('Please enter a valid URL'),
@@ -31,7 +34,10 @@ export function ConnectionForm() {
   const setServerUrl = useServerStore((state) => state.setServerUrl);
   const keepServerRunningOnClose = useServerStore((state) => state.keepServerRunningOnClose);
   const setKeepServerRunningOnClose = useServerStore((state) => state.setKeepServerRunningOnClose);
+  const mode = useServerStore((state) => state.mode);
+  const setMode = useServerStore((state) => state.setMode);
   const { toast } = useToast();
+  const { data: health, isLoading, error: healthError } = useServerHealth();
 
   const form = useForm<ConnectionFormValues>({
     resolver: zodResolver(connectionSchema),
@@ -49,7 +55,7 @@ export function ConnectionForm() {
 
   function onSubmit(data: ConnectionFormValues) {
     setServerUrl(data.serverUrl);
-    form.reset(data); // Reset form state after successful submission
+    form.reset(data);
     toast({
       title: 'Server URL updated',
       description: `Connected to ${data.serverUrl}`,
@@ -57,11 +63,7 @@ export function ConnectionForm() {
   }
 
   return (
-    <Card
-      role="region"
-      aria-label="Server Connection"
-      tabIndex={0}
-    >
+    <Card role="region" aria-label="Server Connection" tabIndex={0}>
       <CardHeader>
         <CardTitle>Server Connection</CardTitle>
       </CardHeader>
@@ -87,6 +89,37 @@ export function ConnectionForm() {
           </form>
         </Form>
 
+        {/* Connection status */}
+        <div className="mt-4">
+          {isLoading ? (
+            <div className="flex items-center gap-2">
+              <Loader2 className="h-4 w-4 animate-spin" />
+              <span className="text-sm text-muted-foreground">Checking connection...</span>
+            </div>
+          ) : healthError ? (
+            <div className="flex items-center gap-2">
+              <XCircle className="h-4 w-4 text-destructive" />
+              <span className="text-sm text-destructive">
+                Connection failed: {healthError.message}
+              </span>
+            </div>
+          ) : health ? (
+            <div className="flex flex-wrap gap-2">
+              <Badge
+                variant={health.model_loaded || health.model_downloaded ? 'default' : 'secondary'}
+              >
+                {health.model_loaded || health.model_downloaded ? 'Model Ready' : 'No Model'}
+              </Badge>
+              <Badge variant={health.gpu_available ? 'default' : 'secondary'}>
+                GPU: {health.gpu_available ? 'Available' : 'Not Available'}
+              </Badge>
+              {health.vram_used_mb && (
+                <Badge variant="outline">VRAM: {health.vram_used_mb.toFixed(0)} MB</Badge>
+              )}
+            </div>
+          ) : null}
+        </div>
+
         <div className="mt-6 pt-6 border-t">
           <div className="flex items-start space-x-3">
             <Checkbox
@@ -119,6 +152,38 @@ export function ConnectionForm() {
             </div>
           </div>
         </div>
+
+        {platform.metadata.isTauri && (
+          <div className="mt-6 pt-6 border-t">
+            <div className="flex items-start space-x-3">
+              <Checkbox
+                id="allowNetworkAccess"
+                checked={mode === 'remote'}
+                onCheckedChange={(checked: boolean) => {
+                  setMode(checked ? 'remote' : 'local');
+                  toast({
+                    title: 'Setting updated',
+                    description: checked
+                      ? 'Network access enabled. Restart the app to apply.'
+                      : 'Network access disabled. Restart the app to apply.',
+                  });
+                }}
+              />
+              <div className="space-y-1">
+                <label
+                  htmlFor="allowNetworkAccess"
+                  className="text-sm font-medium leading-none peer-disabled:cursor-not-allowed peer-disabled:opacity-70 cursor-pointer"
+                >
+                  Allow network access
+                </label>
+                <p className="text-sm text-muted-foreground">
+                  Makes the server accessible from other devices on your network. Restart the app
+                  after changing this setting.
+                </p>
+              </div>
+            </div>
+          </div>
+        )}
       </CardContent>
     </Card>
   );

app/src/components/ServerSettings/GenerationSettings.tsx

@@ -0,0 +1,93 @@
+import { Card, CardContent, CardDescription, CardHeader, CardTitle } from '@/components/ui/card';
+import { Checkbox } from '@/components/ui/checkbox';
+import { Slider } from '@/components/ui/slider';
+import { useServerStore } from '@/stores/serverStore';
+
+export function GenerationSettings() {
+  const maxChunkChars = useServerStore((state) => state.maxChunkChars);
+  const setMaxChunkChars = useServerStore((state) => state.setMaxChunkChars);
+  const crossfadeMs = useServerStore((state) => state.crossfadeMs);
+  const setCrossfadeMs = useServerStore((state) => state.setCrossfadeMs);
+  const normalizeAudio = useServerStore((state) => state.normalizeAudio);
+  const setNormalizeAudio = useServerStore((state) => state.setNormalizeAudio);
+
+  return (
+    <Card role="region" aria-label="Generation Settings" tabIndex={0}>
+      <CardHeader>
+        <CardTitle>Generation Settings</CardTitle>
+        <CardDescription>
+          Controls for long text generation. These settings apply to all engines.
+        </CardDescription>
+      </CardHeader>
+      <CardContent>
+        <div className="space-y-6">
+          <div className="space-y-3">
+            <div className="flex items-center justify-between">
+              <label htmlFor="maxChunkChars" className="text-sm font-medium leading-none">
+                Auto-chunking limit
+              </label>
+              <span className="text-sm tabular-nums text-muted-foreground">
+                {maxChunkChars} chars
+              </span>
+            </div>
+            <Slider
+              id="maxChunkChars"
+              value={[maxChunkChars]}
+              onValueChange={([value]) => setMaxChunkChars(value)}
+              min={100}
+              max={2000}
+              step={50}
+              aria-label="Auto-chunking character limit"
+            />
+            <p className="text-sm text-muted-foreground">
+              Long text is split into chunks at sentence boundaries before generating. Lower values
+              can improve quality for long outputs.
+            </p>
+          </div>
+
+          <div className="space-y-3">
+            <div className="flex items-center justify-between">
+              <label htmlFor="crossfadeMs" className="text-sm font-medium leading-none">
+                Chunk crossfade
+              </label>
+              <span className="text-sm tabular-nums text-muted-foreground">
+                {crossfadeMs === 0 ? 'Cut' : `${crossfadeMs}ms`}
+              </span>
+            </div>
+            <Slider
+              id="crossfadeMs"
+              value={[crossfadeMs]}
+              onValueChange={([value]) => setCrossfadeMs(value)}
+              min={0}
+              max={200}
+              step={10}
+              aria-label="Chunk crossfade duration"
+            />
+            <p className="text-sm text-muted-foreground">
+              Blends audio between chunks to smooth transitions. Set to 0 for a hard cut.
+            </p>
+          </div>
+
+          <div className="flex items-start gap-3">
+            <Checkbox
+              id="normalizeAudio"
+              checked={normalizeAudio}
+              onCheckedChange={setNormalizeAudio}
+            />
+            <div className="space-y-1">
+              <label
+                htmlFor="normalizeAudio"
+                className="text-sm font-medium leading-none cursor-pointer"
+              >
+                Normalize audio
+              </label>
+              <p className="text-sm text-muted-foreground">
+                Adjusts output volume to a consistent level across generations.
+              </p>
+            </div>
+          </div>
+        </div>
+      </CardContent>
+    </Card>
+  );
+}

app/src/components/ServerSettings/GpuAcceleration.tsx

@@ -1,7 +1,6 @@
 import { useQuery, useQueryClient } from '@tanstack/react-query';
-import { AlertCircle, Cpu, Download, Loader2, RotateCw, Trash2, Zap } from 'lucide-react';
+import { AlertCircle, Download, Loader2, RotateCw, Trash2 } from 'lucide-react';
 import { useCallback, useEffect, useRef, useState } from 'react';
-import { Badge } from '@/components/ui/badge';
 import { Button } from '@/components/ui/button';
 import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card';
 import { Progress } from '@/components/ui/progress';
@@ -216,31 +215,19 @@ export function GpuAcceleration() {
   return (
     <Card>
       <CardHeader>
-        <CardTitle className="flex items-center gap-2">
-          <Zap className="h-4 w-4" />
-          GPU Acceleration
-        </CardTitle>
+        <CardTitle>GPU Acceleration</CardTitle>
       </CardHeader>
       <CardContent className="space-y-4">
         {/* Current status */}
-        <div className="flex items-center justify-between">
-          <div className="space-y-1">
-            <div className="text-sm font-medium">Backend</div>
-            <div className="text-sm text-muted-foreground">
-              {isCurrentlyCuda ? 'CUDA (GPU accelerated)' : 'CPU'}
-            </div>
+        <div className="space-y-1">
+          <div className="text-sm font-medium">Backend</div>
+          <div className="text-sm text-muted-foreground">
+            {isCurrentlyCuda
+              ? 'CUDA (GPU accelerated)'
+              : hasNativeGpu
+                ? `${health.backend_type === 'mlx' ? 'MLX' : 'PyTorch'} (GPU accelerated)`
+                : 'CPU'}
           </div>
-          <Badge variant={isCurrentlyCuda ? 'default' : 'secondary'}>
-            {isCurrentlyCuda ? (
-              <>
-                <Zap className="h-3 w-3 mr-1" /> CUDA
-              </>
-            ) : (
-              <>
-                <Cpu className="h-3 w-3 mr-1" /> CPU
-              </>
-            )}
-          </Badge>
         </div>
 
         {/* GPU info from health */}
@@ -257,14 +244,6 @@ export function GpuAcceleration() {
         )}
 
         {/* Native GPU detected - no CUDA download needed */}
-        {hasNativeGpu && (
-          <div className="p-3 rounded-lg bg-accent/10 border border-accent/20">
-            <div className="text-sm">
-              Your system uses <strong>{health.gpu_type}</strong> for acceleration. No additional
-              downloads needed.
-            </div>
-          </div>
-        )}
 
         {/* CUDA download section - only show when native GPU is NOT detected (i.e., Windows/Linux NVIDIA users) */}
         {!hasNativeGpu && (

app/src/components/ServerSettings/ModelManagement.tsx

@@ -7,12 +7,14 @@ import {
   CircleX,
   Download,
   ExternalLink,
+  FolderOpen,
   HardDrive,
   Heart,
   Loader2,
   RotateCcw,
   Scale,
   Trash2,
+  Unplug,
   X,
 } from 'lucide-react';
 import { useCallback, useMemo, useState } from 'react';
@@ -40,6 +42,8 @@ import { useToast } from '@/components/ui/use-toast';
 import { apiClient } from '@/lib/api/client';
 import type { ActiveDownloadTask, HuggingFaceModelInfo, ModelStatus } from '@/lib/api/types';
 import { useModelDownloadToast } from '@/lib/hooks/useModelDownloadToast';
+import { usePlatform } from '@/platform/PlatformContext';
+import { useServerStore } from '@/stores/serverStore';
 
 async function fetchHuggingFaceModelInfo(repoId: string): Promise<HuggingFaceModelInfo> {
   const response = await fetch(`https://huggingface.co/api/models/${repoId}`);
@@ -47,6 +51,29 @@ async function fetchHuggingFaceModelInfo(repoId: string): Promise<HuggingFaceMod
   return response.json();
 }
 
+const MODEL_DESCRIPTIONS: Record<string, string> = {
+  'qwen-tts-1.7B':
+    'High-quality multilingual TTS by Alibaba. Supports 10 languages with natural prosody and voice cloning from short reference audio.',
+  'qwen-tts-0.6B':
+    'Lightweight version of Qwen TTS. Same language support with faster inference, ideal for lower-end hardware.',
+  luxtts:
+    'Lightweight ZipVoice-based TTS designed for high quality voice cloning and 48kHz speech generation at speeds exceeding 150x realtime.',
+  'chatterbox-tts':
+    'Production-grade open source TTS by Resemble AI. Supports 23 languages with voice cloning and emotion exaggeration control.',
+  'chatterbox-turbo':
+    'Streamlined 350M parameter TTS by Resemble AI. High-quality English speech with less compute and VRAM than larger models.',
+  'whisper-base':
+    'Smallest Whisper model (74M parameters). Fast transcription with moderate accuracy.',
+  'whisper-small':
+    'Whisper Small (244M parameters). Good balance of speed and accuracy for transcription.',
+  'whisper-medium':
+    'Whisper Medium (769M parameters). Higher accuracy transcription at moderate speed.',
+  'whisper-large':
+    'Whisper Large (1.5B parameters). Best accuracy for speech-to-text across multiple languages.',
+  'whisper-turbo':
+    'Whisper Large v3 Turbo. Pruned for significantly faster inference while maintaining near-large accuracy.',
+};
+
 function formatDownloads(n: number): string {
   if (n >= 1_000_000) return `${(n / 1_000_000).toFixed(1)}M`;
   if (n >= 1_000) return `${(n / 1_000).toFixed(1)}k`;
@@ -84,6 +111,18 @@ function formatBytes(bytes: number): string {
 export function ModelManagement() {
   const { toast } = useToast();
   const queryClient = useQueryClient();
+  const platform = usePlatform();
+  const customModelsDir = useServerStore((state) => state.customModelsDir);
+  const setCustomModelsDir = useServerStore((state) => state.setCustomModelsDir);
+  const [migrating, setMigrating] = useState(false);
+  const [migrationProgress, setMigrationProgress] = useState<{
+    current: number;
+    total: number;
+    progress: number;
+    filename?: string;
+    status: string;
+  } | null>(null);
+  const [pendingMigrateDir, setPendingMigrateDir] = useState<string | null>(null);
   const [downloadingModel, setDownloadingModel] = useState<string | null>(null);
   const [downloadingDisplayName, setDownloadingDisplayName] = useState<string | null>(null);
   const [consoleOpen, setConsoleOpen] = useState(false);
@@ -103,6 +142,12 @@ export function ModelManagement() {
     refetchInterval: 5000,
   });
 
+  const { data: cacheDir } = useQuery({
+    queryKey: ['modelsCacheDir'],
+    queryFn: () => apiClient.getModelsCacheDir(),
+    staleTime: 1000 * 60 * 5,
+  });
+
   const { data: activeTasks } = useQuery({
     queryKey: ['activeTasks'],
     queryFn: () => apiClient.getActiveTasks(),
@@ -300,6 +345,27 @@ export function ModelManagement() {
     },
   });
 
+  const unloadMutation = useMutation({
+    mutationFn: async (modelName: string) => {
+      return await apiClient.unloadModel(modelName);
+    },
+    onSuccess: async (_data, modelName) => {
+      toast({
+        title: 'Model unloaded',
+        description: `${modelName} has been unloaded from memory.`,
+      });
+      await queryClient.invalidateQueries({ queryKey: ['modelStatus'], refetchType: 'all' });
+      await queryClient.refetchQueries({ queryKey: ['modelStatus'] });
+    },
+    onError: (error: Error) => {
+      toast({
+        title: 'Unload failed',
+        description: error.message,
+        variant: 'destructive',
+      });
+    },
+  });
+
   const formatSize = (sizeMb?: number): string => {
     if (!sizeMb) return 'Unknown size';
     if (sizeMb < 1024) return `${sizeMb.toFixed(1)} MB`;
@@ -320,17 +386,18 @@ export function ModelManagement() {
     setDetailOpen(true);
   };
 
-  const ttsModels = modelStatus?.models.filter((m) => m.model_name.startsWith('qwen-tts')) ?? [];
-  const otherTtsModels =
+  const voiceModels =
     modelStatus?.models.filter(
-      (m) => m.model_name.startsWith('luxtts') || m.model_name.startsWith('chatterbox'),
+      (m) =>
+        m.model_name.startsWith('qwen-tts') ||
+        m.model_name.startsWith('luxtts') ||
+        m.model_name.startsWith('chatterbox'),
     ) ?? [];
   const whisperModels = modelStatus?.models.filter((m) => m.model_name.startsWith('whisper')) ?? [];
 
   // Build sections
   const sections: { label: string; models: ModelStatus[] }[] = [
-    { label: 'Voice Generation', models: ttsModels },
-    ...(otherTtsModels.length > 0 ? [{ label: 'Other Voice Models', models: otherTtsModels }] : []),
+    { label: 'Voice Generation', models: voiceModels },
     { label: 'Transcription', models: whisperModels },
   ];
 
@@ -359,6 +426,87 @@ export function ModelManagement() {
         </p>
       </div>
 
+      {/* Model storage location */}
+      {platform.metadata.isTauri && cacheDir && (
+        <div className="shrink-0 pb-4 border-b mb-4">
+          <div className="flex items-center justify-between gap-2">
+            <div className="min-w-0">
+              <span className="text-xs text-muted-foreground">Storage location</span>
+              <p
+                className="text-xs font-mono text-muted-foreground/70 truncate"
+                title={cacheDir.path}
+              >
+                {cacheDir.path}
+              </p>
+            </div>
+            <div className="flex items-center gap-1 shrink-0">
+              <Button
+                variant="ghost"
+                size="sm"
+                className="text-xs text-muted-foreground h-7 px-2"
+                onClick={async () => {
+                  try {
+                    const { open } = await import('@tauri-apps/plugin-shell');
+                    await open(cacheDir.path);
+                  } catch {
+                    toast({ title: 'Failed to open model folder', variant: 'destructive' });
+                  }
+                }}
+              >
+                <FolderOpen className="h-3 w-3" />
+                Open
+              </Button>
+              <Button
+                variant="ghost"
+                size="sm"
+                className="text-xs text-muted-foreground h-7 px-2"
+                onClick={async () => {
+                  try {
+                    const { open: openDialog } = await import('@tauri-apps/plugin-dialog');
+                    const selected = await openDialog({
+                      directory: true,
+                      title: 'Choose model storage folder',
+                    });
+                    if (!selected) return;
+                    const newDir =
+                      typeof selected === 'string' ? selected : (selected as { path: string }).path;
+                    if (!newDir) return;
+                    setPendingMigrateDir(newDir);
+                  } catch {
+                    toast({ title: 'Failed to open folder picker', variant: 'destructive' });
+                  }
+                }}
+                disabled={migrating}
+              >
+                {migrating ? (
+                  <Loader2 className="h-3 w-3 animate-spin" />
+                ) : (
+                  <FolderOpen className="h-3 w-3" />
+                )}
+                {migrating ? 'Migrating...' : 'Change'}
+              </Button>
+              {customModelsDir && (
+                <Button
+                  variant="ghost"
+                  size="sm"
+                  className="text-xs text-muted-foreground h-7 px-2"
+                  disabled={migrating}
+                  onClick={async () => {
+                    setCustomModelsDir(null);
+                    toast({ title: 'Reset to default location. Restarting server...' });
+                    await platform.lifecycle.restartServer('');
+                    queryClient.invalidateQueries();
+                  }}
+                >
+                  <RotateCcw className="h-3 w-3" />
+                  Reset
+                </Button>
+              )}
+            </div>
+          </div>
+        </div>
+      )}
+
       {/* Model list */}
       {isLoading ? (
         <div className="flex items-center justify-center py-16">
@@ -434,9 +582,7 @@ export function ModelManagement() {
                             {formatSize(model.size_mb)}
                           </span>
                         )}
-                        {!model.downloaded && !isDownloading && !hasError && (
-                          <span className="text-xs text-muted-foreground/60">Not downloaded</span>
-                        )}
+
                         <ChevronRight className="h-4 w-4 text-muted-foreground/40 group-hover:text-muted-foreground transition-colors" />
                       </div>
                     </button>
@@ -542,25 +688,12 @@ export function ModelManagement() {
                       Loaded
                     </Badge>
                   )}
-                  {freshSelectedModel.downloaded && !freshSelectedModel.loaded && (
-                    <Badge variant="secondary" className="text-xs">
-                      <CircleCheck className="h-3 w-3 mr-1" />
-                      Downloaded
-                    </Badge>
-                  )}
                   {selectedState?.hasError && (
                     <Badge variant="destructive" className="text-xs">
                       <CircleX className="h-3 w-3 mr-1" />
                       Error
                     </Badge>
                   )}
-                  {!freshSelectedModel.downloaded &&
-                    !selectedState?.isDownloading &&
-                    !selectedState?.hasError && (
-                      <Badge variant="outline" className="text-xs text-muted-foreground">
-                        Not downloaded
-                      </Badge>
-                    )}
                 </div>
 
                 {/* HuggingFace model card info */}
@@ -571,26 +704,15 @@ export function ModelManagement() {
                   </div>
                 )}
 
+                {/* Description */}
+                {MODEL_DESCRIPTIONS[freshSelectedModel.model_name] && (
+                  <p className="text-xs text-muted-foreground leading-relaxed">
+                    {MODEL_DESCRIPTIONS[freshSelectedModel.model_name]}
+                  </p>
+                )}
+
                 {hfModelInfo && (
                   <div className="space-y-3">
-                    {/* Stats row */}
-                    <div className="flex items-center gap-4 text-xs text-muted-foreground">
-                      <span className="flex items-center gap-1" title="Downloads">
-                        <Download className="h-3.5 w-3.5" />
-                        {formatDownloads(hfModelInfo.downloads)}
-                      </span>
-                      <span className="flex items-center gap-1" title="Likes">
-                        <Heart className="h-3.5 w-3.5" />
-                        {formatDownloads(hfModelInfo.likes)}
-                      </span>
-                      {license && (
-                        <span className="flex items-center gap-1" title="License">
-                          <Scale className="h-3.5 w-3.5" />
-                          {formatLicense(license)}
-                        </span>
-                      )}
-                    </div>
-
                     {/* Pipeline tag + author */}
                     <div className="flex flex-wrap gap-1.5">
                       {hfModelInfo.pipeline_tag && (
@@ -610,6 +732,24 @@ export function ModelManagement() {
                       )}
                     </div>
 
+                    {/* Stats row */}
+                    <div className="flex items-center gap-4 text-xs text-muted-foreground">
+                      <span className="flex items-center gap-1" title="Downloads">
+                        <Download className="h-3.5 w-3.5" />
+                        {formatDownloads(hfModelInfo.downloads)}
+                      </span>
+                      <span className="flex items-center gap-1" title="Likes">
+                        <Heart className="h-3.5 w-3.5" />
+                        {formatDownloads(hfModelInfo.likes)}
+                      </span>
+                      {license && (
+                        <span className="flex items-center gap-1" title="License">
+                          <Scale className="h-3.5 w-3.5" />
+                          {formatLicense(license)}
+                        </span>
+                      )}
+                    </div>
+
                     {/* Languages */}
                     {hfModelInfo.cardData?.language && hfModelInfo.cardData.language.length > 0 && (
                       <div>
@@ -625,8 +765,8 @@ export function ModelManagement() {
 
                 {/* Disk size */}
                 {freshSelectedModel.downloaded && freshSelectedModel.size_mb && (
-                  <div className="flex items-center gap-2 text-sm text-muted-foreground">
-                    <HardDrive className="h-4 w-4" />
+                  <div className="flex items-center gap-2 text-xs text-muted-foreground">
+                    <HardDrive className="h-3.5 w-3.5" />
                     <span>{formatSize(freshSelectedModel.size_mb)} on disk</span>
                   </div>
                 )}
@@ -639,7 +779,7 @@ export function ModelManagement() {
                 )}
 
                 {/* Actions */}
-                <div className="flex items-center gap-2 pt-2 border-t">
+                <div className="flex items-center gap-2 pt-2">
                   {selectedState?.hasError ? (
                     <>
                       <Button
@@ -697,26 +837,46 @@ export function ModelManagement() {
                       </Button>
                     </>
                   ) : freshSelectedModel.downloaded ? (
-                    <Button
-                      size="sm"
-                      onClick={() => {
-                        setModelToDelete({
-                          name: freshSelectedModel.model_name,
-                          displayName: freshSelectedModel.display_name,
-                          sizeMb: freshSelectedModel.size_mb,
-                        });
-                        setDeleteDialogOpen(true);
-                      }}
-                      variant="outline"
-                      disabled={freshSelectedModel.loaded}
-                      title={
-                        freshSelectedModel.loaded ? 'Unload model before deleting' : 'Delete model'
-                      }
-                      className="flex-1"
-                    >
-                      <Trash2 className="h-4 w-4 mr-2" />
-                      {freshSelectedModel.loaded ? 'Unload to Delete' : 'Delete Model'}
-                    </Button>
+                    <div className="flex gap-2 flex-1">
+                      {freshSelectedModel.loaded && (
+                        <Button
+                          size="sm"
+                          onClick={() => unloadMutation.mutate(freshSelectedModel.model_name)}
+                          variant="outline"
+                          disabled={unloadMutation.isPending}
+                          className="flex-1"
+                        >
+                          {unloadMutation.isPending ? (
+                            <Loader2 className="h-4 w-4 mr-2 animate-spin" />
+                          ) : (
+                            <Unplug className="h-4 w-4 mr-2" />
+                          )}
+                          {unloadMutation.isPending ? 'Unloading...' : 'Unload'}
+                        </Button>
+                      )}
+                      <Button
+                        size="sm"
+                        onClick={() => {
+                          setModelToDelete({
+                            name: freshSelectedModel.model_name,
+                            displayName: freshSelectedModel.display_name,
+                            sizeMb: freshSelectedModel.size_mb,
+                          });
+                          setDeleteDialogOpen(true);
+                        }}
+                        variant="outline"
+                        disabled={freshSelectedModel.loaded}
+                        title={
+                          freshSelectedModel.loaded
+                            ? 'Unload model before deleting'
+                            : 'Delete model'
+                        }
+                        className="flex-1"
+                      >
+                        <Trash2 className="h-4 w-4 mr-2" />
+                        Delete Model
+                      </Button>
+                    </div>
                   ) : (
                     <Button
                       size="sm"
@@ -773,7 +933,127 @@ export function ModelManagement() {
           </AlertDialogFooter>
         </AlertDialogContent>
       </AlertDialog>
-    </Card>
+
+      {/* Migration confirmation dialog */}
+      <AlertDialog
+        open={!!pendingMigrateDir}
+        onOpenChange={(open) => !open && setPendingMigrateDir(null)}
+      >
+        <AlertDialogContent>
+          <AlertDialogHeader>
+            <AlertDialogTitle>Move models to new location?</AlertDialogTitle>
+            <AlertDialogDescription>
+              The server will shut down while models are being moved to the new folder. It will
+              restart automatically once the migration is complete.
+            </AlertDialogDescription>
+          </AlertDialogHeader>
+          <div
+            className="text-xs font-mono text-muted-foreground bg-muted/50 rounded px-3 py-2 truncate"
+            title={pendingMigrateDir ?? ''}
+          >
+            {pendingMigrateDir}
+          </div>
+          <AlertDialogFooter>
+            <AlertDialogCancel>Cancel</AlertDialogCancel>
+            <AlertDialogAction
+              onClick={async () => {
+                if (!pendingMigrateDir) return;
+                const newDir = pendingMigrateDir;
+                setPendingMigrateDir(null);
+                setMigrating(true);
+                setMigrationProgress({
+                  current: 0,
+                  total: 0,
+                  progress: 0,
+                  status: 'downloading',
+                  filename: 'Preparing...',
+                });
+                try {
+                  // Start the migration (background task)
+                  await apiClient.migrateModels(newDir);
+
+                  // Connect to SSE for progress
+                  await new Promise<void>((resolve, reject) => {
+                    const es = new EventSource(apiClient.getMigrationProgressUrl());
+                    es.onmessage = (event) => {
+                      try {
+                        const data = JSON.parse(event.data);
+                        setMigrationProgress(data);
+                        if (data.status === 'complete') {
+                          es.close();
+                          resolve();
+                        } else if (data.status === 'error') {
+                          es.close();
+                          reject(new Error(data.error || 'Migration failed'));
+                        }
+                      } catch {
+                        /* ignore parse errors */
+                      }
+                    };
+                    es.onerror = () => {
+                      es.close();
+                      reject(new Error('Lost connection during migration'));
+                    };
+                  });
+
+                  setCustomModelsDir(newDir);
+                  setMigrationProgress({
+                    current: 1,
+                    total: 1,
+                    progress: 100,
+                    status: 'complete',
+                    filename: 'Restarting server...',
+                  });
+                  await platform.lifecycle.restartServer(newDir);
+                  queryClient.invalidateQueries();
+                  toast({ title: 'Models moved successfully' });
+                } catch (e) {
+                  toast({
+                    title: 'Migration failed',
+                    description: e instanceof Error ? e.message : 'Failed to migrate models',
+                    variant: 'destructive',
+                  });
+                } finally {
+                  setMigrating(false);
+                  setMigrationProgress(null);
+                }
+              }}
+            >
+              Move Models
+            </AlertDialogAction>
+          </AlertDialogFooter>
+        </AlertDialogContent>
+      </AlertDialog>
+
+      {/* Migration progress overlay */}
+      {migrating && migrationProgress && (
+        <div className="fixed inset-0 z-50 bg-background/95 backdrop-blur-sm flex items-center justify-center">
+          <div className="w-full max-w-md px-8 space-y-6 text-center">
+            <div className="space-y-2">
+              <Loader2 className="h-8 w-8 animate-spin mx-auto text-muted-foreground" />
+              <h2 className="text-lg font-semibold">Moving models</h2>
+              <p className="text-sm text-muted-foreground">
+                {migrationProgress.status === 'complete'
+                  ? 'Restarting server...'
+                  : 'The server is offline while models are being moved.'}
+              </p>
+            </div>
+            {migrationProgress.total > 0 && (
+              <div className="space-y-2">
+                <Progress value={migrationProgress.progress} className="h-2" />
+                <div className="flex justify-between text-xs text-muted-foreground">
+                  <span className="truncate max-w-[60%]">{migrationProgress.filename}</span>
+                  <span>
+                    {formatBytes(migrationProgress.current)} /{' '}
+                    {formatBytes(migrationProgress.total)}
+                  </span>
+                </div>
+              </div>
+            )}
+          </div>
+        </div>
+      )}
+    </div>
   );
 }
 
@@ -782,13 +1062,13 @@ interface ModelItemProps {
     model_name: string;
     display_name: string;
     downloaded: boolean;
-    downloading?: boolean;  // From server - true if download in progress
+    downloading?: boolean; // From server - true if download in progress
     size_mb?: number;
     loaded: boolean;
   };
   onDownload: () => void;
   onDelete: () => void;
-  isDownloading: boolean;  // Local state - true if user just clicked download
+  isDownloading: boolean; // Local state - true if user just clicked download
   formatSize: (sizeMb?: number) => string;
 }
 
@@ -848,16 +1128,19 @@ function ModelItem({ model, onDownload, onDelete, isDownloading, formatSize }: M
               disabled={model.loaded}
               title={model.loaded ? 'Unload model before deleting' : 'Delete model'}
               aria-label={
-                model.loaded
-                  ? 'Unload model before deleting'
-                  : `Delete ${model.display_name}`
+                model.loaded ? 'Unload model before deleting' : `Delete ${model.display_name}`
               }
             >
               <Trash2 className="h-4 w-4" />
             </Button>
           </div>
         ) : showDownloading ? (
-          <Button size="sm" variant="outline" disabled aria-label={`${model.display_name} downloading`}>
+          <Button
+            size="sm"
+            variant="outline"
+            disabled
+            aria-label={`${model.display_name} downloading`}
+          >
             <Loader2 className="h-4 w-4 mr-2 animate-spin" />
             Downloading...
           </Button>

app/src/components/ServerSettings/ServerStatus.tsx

@@ -3,18 +3,13 @@ import { Badge } from '@/components/ui/badge';
 import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card';
 import { useServerHealth } from '@/lib/hooks/useServer';
 import { useServerStore } from '@/stores/serverStore';
-import { ModelProgress } from './ModelProgress';
 
 export function ServerStatus() {
   const { data: health, isLoading, error } = useServerHealth();
   const serverUrl = useServerStore((state) => state.serverUrl);
 
   return (
-    <Card
-      role="region"
-      aria-label="Server Status"
-      tabIndex={0}
-    >
+    <Card role="region" aria-label="Server Status" tabIndex={0}>
       <CardHeader>
         <CardTitle>Server Status</CardTitle>
       </CardHeader>
@@ -24,16 +19,6 @@ export function ServerStatus() {
           <div className="font-mono text-sm">{serverUrl}</div>
         </div>
 
-        {/* Model download progress */}
-        <div className="space-y-2">
-          <ModelProgress modelName="qwen-tts-1.7B" displayName="Qwen TTS 1.7B" />
-          <ModelProgress modelName="qwen-tts-0.6B" displayName="Qwen TTS 0.6B" />
-          <ModelProgress modelName="whisper-base" displayName="Whisper Base" />
-          <ModelProgress modelName="whisper-small" displayName="Whisper Small" />
-          <ModelProgress modelName="whisper-medium" displayName="Whisper Medium" />
-          <ModelProgress modelName="whisper-large" displayName="Whisper Large" />
-        </div>
-
         {isLoading ? (
           <div className="flex items-center gap-2">
             <Loader2 className="h-4 w-4 animate-spin" />

app/src/components/ServerTab/ServerTab.tsx

@@ -1,19 +1,25 @@
 import { ConnectionForm } from '@/components/ServerSettings/ConnectionForm';
+import { GenerationSettings } from '@/components/ServerSettings/GenerationSettings';
 import { GpuAcceleration } from '@/components/ServerSettings/GpuAcceleration';
-import { ServerStatus } from '@/components/ServerSettings/ServerStatus';
 import { UpdateStatus } from '@/components/ServerSettings/UpdateStatus';
+import { BOTTOM_SAFE_AREA_PADDING } from '@/lib/constants/ui';
+import { cn } from '@/lib/utils/cn';
 import { usePlatform } from '@/platform/PlatformContext';
+import { usePlayerStore } from '@/stores/playerStore';
 
 export function ServerTab() {
   const platform = usePlatform();
+  const isPlayerVisible = !!usePlayerStore((state) => state.audioUrl);
   return (
-    <div className="space-y-4 overflow-y-auto flex flex-col">
+    <div
+      className={cn('overflow-y-auto flex flex-col', isPlayerVisible && BOTTOM_SAFE_AREA_PADDING)}
+    >
       <div className="grid gap-4 md:grid-cols-2">
         <ConnectionForm />
-        <ServerStatus />
+        <GenerationSettings />
+        {platform.metadata.isTauri && <GpuAcceleration />}
+        {platform.metadata.isTauri && <UpdateStatus />}
       </div>
-      {platform.metadata.isTauri && <GpuAcceleration />}
-      {platform.metadata.isTauri && <UpdateStatus />}
       <div className="py-8 text-center text-sm text-muted-foreground">
         Created by{' '}
         <a

app/src/lib/api/client.ts

@@ -316,6 +316,21 @@ class ApiClient {
     return this.request<ModelStatusListResponse>('/models/status');
   }
 
+  async getModelsCacheDir(): Promise<{ path: string }> {
+    return this.request<{ path: string }>('/models/cache-dir');
+  }
+
+  async migrateModels(destination: string): Promise<{ source: string; destination: string }> {
+    return this.request('/models/migrate', {
+      method: 'POST',
+      body: JSON.stringify({ destination }),
+    });
+  }
+
+  getMigrationProgressUrl(): string {
+    return `${this.getBaseUrl()}/models/migrate/progress`;
+  }
+
   async triggerModelDownload(modelName: string): Promise<{ message: string }> {
     console.log(
       '[API] triggerModelDownload called for:',
@@ -337,6 +352,12 @@ class ApiClient {
     });
   }
 
+  async unloadModel(modelName: string): Promise<{ message: string }> {
+    return this.request<{ message: string }>(`/models/${modelName}/unload`, {
+      method: 'POST',
+    });
+  }
+
   async cancelDownload(modelName: string): Promise<{ message: string }> {
     return this.request<{ message: string }>('/models/download/cancel', {
       method: 'POST',

app/src/lib/api/types.ts

@@ -34,8 +34,11 @@ export interface GenerationRequest {
   language: LanguageCode;
   seed?: number;
   model_size?: '1.7B' | '0.6B';
-  engine?: 'qwen' | 'luxtts' | 'chatterbox';
+  engine?: 'qwen' | 'luxtts' | 'chatterbox' | 'chatterbox_turbo';
   instruct?: string;
+  max_chunk_chars?: number;
+  crossfade_ms?: number;
+  normalize?: boolean;
 }
 
 export interface GenerationResponse {

app/src/lib/constants/languages.ts

@@ -1,27 +1,86 @@
 /**
- * Supported languages for voice generation.
- * Most languages use Qwen3-TTS; Hebrew uses Chatterbox TTS.
+ * Supported languages for voice generation, per engine.
+ *
+ * Qwen3-TTS supports 10 languages.
+ * LuxTTS is English-only.
+ * Chatterbox Multilingual supports 23 languages.
+ * Chatterbox Turbo is English-only.
  */
 
-export const SUPPORTED_LANGUAGES = {
-  zh: 'Chinese',
+/** All languages that any engine supports. */
+export const ALL_LANGUAGES = {
+  ar: 'Arabic',
+  da: 'Danish',
+  de: 'German',
+  el: 'Greek',
   en: 'English',
+  es: 'Spanish',
+  fi: 'Finnish',
+  fr: 'French',
+  he: 'Hebrew',
+  hi: 'Hindi',
+  it: 'Italian',
   ja: 'Japanese',
   ko: 'Korean',
-  de: 'German',
-  fr: 'French',
-  ru: 'Russian',
+  ms: 'Malay',
+  nl: 'Dutch',
+  no: 'Norwegian',
+  pl: 'Polish',
   pt: 'Portuguese',
-  es: 'Spanish',
-  it: 'Italian',
-  he: 'Hebrew',
+  ru: 'Russian',
+  sv: 'Swedish',
+  sw: 'Swahili',
+  tr: 'Turkish',
+  zh: 'Chinese',
 } as const;
 
-export type LanguageCode = keyof typeof SUPPORTED_LANGUAGES;
+export type LanguageCode = keyof typeof ALL_LANGUAGES;
 
-export const LANGUAGE_CODES = Object.keys(SUPPORTED_LANGUAGES) as LanguageCode[];
+/** Per-engine supported language codes. */
+export const ENGINE_LANGUAGES: Record<string, readonly LanguageCode[]> = {
+  qwen: ['zh', 'en', 'ja', 'ko', 'de', 'fr', 'ru', 'pt', 'es', 'it'],
+  luxtts: ['en'],
+  chatterbox: [
+    'ar',
+    'da',
+    'de',
+    'el',
+    'en',
+    'es',
+    'fi',
+    'fr',
+    'he',
+    'hi',
+    'it',
+    'ja',
+    'ko',
+    'ms',
+    'nl',
+    'no',
+    'pl',
+    'pt',
+    'ru',
+    'sv',
+    'sw',
+    'tr',
+    'zh',
+  ],
+  chatterbox_turbo: ['en'],
+} as const;
 
+/** Helper: get language options for a given engine. */
+export function getLanguageOptionsForEngine(engine: string) {
+  const codes = ENGINE_LANGUAGES[engine] ?? ENGINE_LANGUAGES.qwen;
+  return codes.map((code) => ({
+    value: code,
+    label: ALL_LANGUAGES[code],
+  }));
+}
+
+// ── Backwards-compatible exports used elsewhere ──────────────────────
+export const SUPPORTED_LANGUAGES = ALL_LANGUAGES;
+export const LANGUAGE_CODES = Object.keys(ALL_LANGUAGES) as LanguageCode[];
 export const LANGUAGE_OPTIONS = LANGUAGE_CODES.map((code) => ({
   value: code,
-  label: SUPPORTED_LANGUAGES[code],
+  label: ALL_LANGUAGES[code],
 }));

app/src/lib/hooks/useGenerationForm.ts

@@ -9,14 +9,15 @@ import { useGeneration } from '@/lib/hooks/useGeneration';
 import { useModelDownloadToast } from '@/lib/hooks/useModelDownloadToast';
 import { useGenerationStore } from '@/stores/generationStore';
 import { usePlayerStore } from '@/stores/playerStore';
+import { useServerStore } from '@/stores/serverStore';
 
 const generationSchema = z.object({
-  text: z.string().min(1, 'Text is required').max(5000),
+  text: z.string().min(1, 'Text is required').max(50000),
   language: z.enum(LANGUAGE_CODES as [LanguageCode, ...LanguageCode[]]),
   seed: z.number().int().optional(),
   modelSize: z.enum(['1.7B', '0.6B']).optional(),
   instruct: z.string().max(500).optional(),
-  engine: z.enum(['qwen', 'luxtts', 'chatterbox']).optional(),
+  engine: z.enum(['qwen', 'luxtts', 'chatterbox', 'chatterbox_turbo']).optional(),
 });
 
 export type GenerationFormValues = z.infer<typeof generationSchema>;
@@ -31,6 +32,9 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
   const generation = useGeneration();
   const setAudioWithAutoPlay = usePlayerStore((state) => state.setAudioWithAutoPlay);
   const setIsGenerating = useGenerationStore((state) => state.setIsGenerating);
+  const maxChunkChars = useServerStore((state) => state.maxChunkChars);
+  const crossfadeMs = useServerStore((state) => state.crossfadeMs);
+  const normalizeAudio = useServerStore((state) => state.normalizeAudio);
   const [downloadingModelName, setDownloadingModelName] = useState<string | null>(null);
   const [downloadingDisplayName, setDownloadingDisplayName] = useState<string | null>(null);
 
@@ -75,15 +79,19 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
           ? 'luxtts'
           : engine === 'chatterbox'
             ? 'chatterbox-tts'
-            : `qwen-tts-${data.modelSize}`;
+            : engine === 'chatterbox_turbo'
+              ? 'chatterbox-turbo'
+              : `qwen-tts-${data.modelSize}`;
       const displayName =
         engine === 'luxtts'
           ? 'LuxTTS'
           : engine === 'chatterbox'
             ? 'Chatterbox TTS'
-            : data.modelSize === '1.7B'
-              ? 'Qwen TTS 1.7B'
-              : 'Qwen TTS 0.6B';
+            : engine === 'chatterbox_turbo'
+              ? 'Chatterbox Turbo'
+              : data.modelSize === '1.7B'
+                ? 'Qwen TTS 1.7B'
+                : 'Qwen TTS 0.6B';
 
       try {
         const modelStatus = await apiClient.getModelStatus();
@@ -106,6 +114,9 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
         model_size: isQwen ? data.modelSize : undefined,
         engine,
         instruct: isQwen ? data.instruct || undefined : undefined,
+        max_chunk_chars: maxChunkChars,
+        crossfade_ms: crossfadeMs,
+        normalize: normalizeAudio,
       });
 
       toast({

app/src/lib/hooks/useRestoreActiveTasks.tsx

@@ -1,23 +1,23 @@
 import { useCallback, useEffect, useRef, useState } from 'react';
 import { apiClient } from '@/lib/api/client';
-import { useGenerationStore } from '@/stores/generationStore';
 import type { ActiveDownloadTask } from '@/lib/api/types';
+import { useGenerationStore } from '@/stores/generationStore';
 
 // Polling interval in milliseconds
-const POLL_INTERVAL = 2000;
+const POLL_INTERVAL = 30000;
 
 /**
  * Hook to monitor active tasks (downloads and generations).
  * Polls the server periodically to catch downloads triggered from anywhere
  * (transcription, generation, explicit download, etc.).
- * 
+ *
  * Returns the active downloads so components can render download toasts.
  */
 export function useRestoreActiveTasks() {
   const [activeDownloads, setActiveDownloads] = useState<ActiveDownloadTask[]>([]);
   const setIsGenerating = useGenerationStore((state) => state.setIsGenerating);
   const setActiveGenerationId = useGenerationStore((state) => state.setActiveGenerationId);
-  
+
   // Track which downloads we've seen to detect new ones
   const seenDownloadsRef = useRef<Set<string>>(new Set());
 
@@ -41,14 +41,14 @@ export function useRestoreActiveTasks() {
       // Update active downloads
       // Keep track of all active downloads (including new ones)
       const currentDownloadNames = new Set(tasks.downloads.map((d) => d.model_name));
-      
+
       // Remove completed downloads from our seen set
       for (const name of seenDownloadsRef.current) {
         if (!currentDownloadNames.has(name)) {
           seenDownloadsRef.current.delete(name);
         }
       }
-      
+
       // Add new downloads to seen set
       for (const download of tasks.downloads) {
         seenDownloadsRef.current.add(download.model_name);

app/src/platform/types.ts

@@ -49,9 +49,9 @@ export interface PlatformAudio {
 }
 
 export interface PlatformLifecycle {
-  startServer(remote?: boolean): Promise<string>;
+  startServer(remote?: boolean, modelsDir?: string | null): Promise<string>;
   stopServer(): Promise<void>;
-  restartServer(): Promise<string>;
+  restartServer(modelsDir?: string | null): Promise<string>;
   setKeepServerRunning(keep: boolean): Promise<void>;
   setupWindowCloseHandler(): Promise<void>;
   onServerReady?: () => void;

app/src/stores/serverStore.ts

@@ -13,6 +13,18 @@ interface ServerStore {
 
   keepServerRunningOnClose: boolean;
   setKeepServerRunningOnClose: (keepRunning: boolean) => void;
+
+  maxChunkChars: number;
+  setMaxChunkChars: (value: number) => void;
+
+  crossfadeMs: number;
+  setCrossfadeMs: (value: number) => void;
+
+  normalizeAudio: boolean;
+  setNormalizeAudio: (value: boolean) => void;
+
+  customModelsDir: string | null;
+  setCustomModelsDir: (dir: string | null) => void;
 }
 
 export const useServerStore = create<ServerStore>()(
@@ -29,6 +41,18 @@ export const useServerStore = create<ServerStore>()(
 
       keepServerRunningOnClose: false,
       setKeepServerRunningOnClose: (keepRunning) => set({ keepServerRunningOnClose: keepRunning }),
+
+      maxChunkChars: 800,
+      setMaxChunkChars: (value) => set({ maxChunkChars: value }),
+
+      crossfadeMs: 50,
+      setCrossfadeMs: (value) => set({ crossfadeMs: value }),
+
+      normalizeAudio: true,
+      setNormalizeAudio: (value) => set({ normalizeAudio: value }),
+
+      customModelsDir: null,
+      setCustomModelsDir: (dir) => set({ customModelsDir: dir }),
     }),
     {
       name: 'voicebox-server',

backend/backends/__init__.py

@@ -122,6 +122,7 @@ TTS_ENGINES = {
     "qwen": "Qwen TTS",
     "luxtts": "LuxTTS",
     "chatterbox": "Chatterbox TTS",
+    "chatterbox_turbo": "Chatterbox Turbo",
 }
 
 
@@ -171,6 +172,9 @@ def get_tts_backend_for_engine(engine: str) -> TTSBackend:
         elif engine == "chatterbox":
             from .chatterbox_backend import ChatterboxTTSBackend
             backend = ChatterboxTTSBackend()
+        elif engine == "chatterbox_turbo":
+            from .chatterbox_turbo_backend import ChatterboxTurboTTSBackend
+            backend = ChatterboxTurboTTSBackend()
         else:
             raise ValueError(f"Unknown TTS engine: {engine}. Supported: {list(TTS_ENGINES.keys())}")
         

backend/backends/chatterbox_backend.py

@@ -136,6 +136,10 @@ class ChatterboxTTSBackend:
             import torch
             from chatterbox.mtl_tts import ChatterboxMultilingualTTS
 
+            # Load into a local variable first, apply all patches, then
+            # assign to self.model.  This avoids leaving a half-initialised
+            # model on self.model if any patch step raises an exception.
+            #
             # Monkey-patch torch.load for CPU loading. The model's .pt files
             # were saved on CUDA; from_pretrained() doesn't pass map_location
             # so loading on CPU fails without this.
@@ -150,13 +154,13 @@ class ChatterboxTTSBackend:
                     with ChatterboxTTSBackend._load_lock:
                         torch.load = _patched_load
                         try:
-                            self.model = ChatterboxMultilingualTTS.from_pretrained(
+                            model = ChatterboxMultilingualTTS.from_pretrained(
                                 device=device,
                             )
                         finally:
                             torch.load = _orig_torch_load
                 else:
-                    self.model = ChatterboxMultilingualTTS.from_pretrained(
+                    model = ChatterboxMultilingualTTS.from_pretrained(
                         device=device,
                     )
             finally:
@@ -165,7 +169,7 @@ class ChatterboxTTSBackend:
             # Fix: transformers >= 4.36 defaults LlamaModel to sdpa attention
             # which doesn't support output_attentions=True (needed by
             # Chatterbox's AlignmentStreamAnalyzer). Force eager attention.
-            t3_tfmr = self.model.t3.tfmr
+            t3_tfmr = model.t3.tfmr
             if hasattr(t3_tfmr, "config") and hasattr(
                 t3_tfmr.config, "_attn_implementation"
             ):
@@ -178,6 +182,36 @@ class ChatterboxTTSBackend:
                 progress_manager.mark_complete(model_name)
                 task_manager.complete_download(model_name)
 
+            # Patch float64 → float32 dtype mismatches in upstream chatterbox.
+            # librosa.load returns float64 numpy; multiple upstream code paths
+            # convert it to a torch tensor via torch.from_numpy() without
+            # casting, then matmul it against float32 model weights.
+            import types
+
+            # Patch S3Tokenizer (used by s3gen.tokenizer)
+            _tokzr = model.s3gen.tokenizer
+            _orig_log_mel = _tokzr.log_mel_spectrogram.__func__
+
+            def _f32_log_mel(self_tokzr, audio, padding=0):
+                import torch as _torch
+                if _torch.is_tensor(audio):
+                    audio = audio.float()
+                return _orig_log_mel(self_tokzr, audio, padding)
+
+            _tokzr.log_mel_spectrogram = types.MethodType(_f32_log_mel, _tokzr)
+
+            # Patch VoiceEncoder
+            _ve = model.ve
+            _orig_ve_forward = _ve.forward.__func__
+
+            def _f32_ve_forward(self_ve, mels):
+                return _orig_ve_forward(self_ve, mels.float())
+
+            _ve.forward = types.MethodType(_f32_ve_forward, _ve)
+
+            # All patches applied successfully — publish the model
+            self.model = model
+
             logger.info("Chatterbox Multilingual TTS loaded successfully")
 
         except ImportError as e:

backend/backends/chatterbox_turbo_backend.py

@@ -0,0 +1,345 @@
+"""
+Chatterbox Turbo TTS backend implementation.
+
+Wraps ChatterboxTurboTTS from chatterbox-tts for fast, English-only
+voice cloning with paralinguistic tag support ([laugh], [cough], etc.).
+Forces CPU on macOS due to known MPS tensor issues.
+"""
+
+import asyncio
+import logging
+import platform
+import threading
+from pathlib import Path
+from typing import ClassVar, List, Optional, Tuple
+
+import numpy as np
+
+from . import TTSBackend
+from ..utils.audio import normalize_audio, load_audio
+from ..utils.progress import get_progress_manager
+from ..utils.tasks import get_task_manager
+
+logger = logging.getLogger(__name__)
+
+CHATTERBOX_TURBO_HF_REPO = "ResembleAI/chatterbox-turbo"
+
+# Files that must be present for the turbo model
+_TURBO_WEIGHT_FILES = [
+    "t3_turbo_v1.safetensors",
+    "s3gen_meanflow.safetensors",
+    "ve.safetensors",
+]
+
+
+class ChatterboxTurboTTSBackend:
+    """Chatterbox Turbo TTS backend — fast, English-only, with paralinguistic tags."""
+
+    # Class-level lock for torch.load monkey-patching
+    _load_lock: ClassVar[threading.Lock] = threading.Lock()
+
+    def __init__(self):
+        self.model = None
+        self.model_size = "default"
+        self._device = None
+        self._model_load_lock = asyncio.Lock()
+
+    def _get_device(self) -> str:
+        """Get the best available device. Forces CPU on macOS (MPS issue)."""
+        if platform.system() == "Darwin":
+            return "cpu"
+        try:
+            import torch
+
+            if torch.cuda.is_available():
+                return "cuda"
+        except ImportError:
+            pass
+        return "cpu"
+
+    def is_loaded(self) -> bool:
+        return self.model is not None
+
+    def _get_model_path(self, model_size: str = "default") -> str:
+        return CHATTERBOX_TURBO_HF_REPO
+
+    def _is_model_cached(self, model_size: str = "default") -> bool:
+        """Check if the Chatterbox Turbo model is cached locally."""
+        try:
+            from huggingface_hub import constants as hf_constants
+
+            repo_cache = Path(hf_constants.HF_HUB_CACHE) / (
+                "models--" + CHATTERBOX_TURBO_HF_REPO.replace("/", "--")
+            )
+
+            if not repo_cache.exists():
+                return False
+
+            blobs_dir = repo_cache / "blobs"
+            if blobs_dir.exists() and any(blobs_dir.glob("*.incomplete")):
+                return False
+
+            # Check for turbo weight files
+            snapshots_dir = repo_cache / "snapshots"
+            if snapshots_dir.exists():
+                for fname in _TURBO_WEIGHT_FILES:
+                    if not any(snapshots_dir.rglob(fname)):
+                        return False
+                return True
+
+            return False
+        except Exception as e:
+            logger.warning(f"Error checking Chatterbox Turbo cache: {e}")
+            return False
+
+    async def load_model(self, model_size: str = "default") -> None:
+        """Load the Chatterbox Turbo model."""
+        if self.model is not None:
+            return
+        async with self._model_load_lock:
+            if self.model is not None:
+                return
+            await asyncio.to_thread(self._load_model_sync)
+
+    def _load_model_sync(self):
+        """Synchronous model loading."""
+        from ..utils.hf_progress import HFProgressTracker, create_hf_progress_callback
+
+        progress_manager = get_progress_manager()
+        task_manager = get_task_manager()
+        model_name = "chatterbox-turbo"
+
+        is_cached = self._is_model_cached()
+
+        # Set up HF progress tracking (intercepts tqdm for file-level progress)
+        progress_callback = create_hf_progress_callback(model_name, progress_manager)
+        tracker = HFProgressTracker(progress_callback, filter_non_downloads=is_cached)
+        tracker_context = tracker.patch_download()
+        tracker_context.__enter__()
+
+        if not is_cached:
+            task_manager.start_download(model_name)
+            progress_manager.update_progress(
+                model_name=model_name,
+                current=0,
+                total=0,
+                filename="Connecting to HuggingFace...",
+                status="downloading",
+            )
+
+        try:
+            device = self._get_device()
+            self._device = device
+
+            logger.info(f"Loading Chatterbox Turbo TTS on {device}...")
+
+            import torch
+            from huggingface_hub import snapshot_download
+            from chatterbox.tts_turbo import ChatterboxTurboTTS
+
+            # Download model files ourselves so we can pass token=None
+            # (upstream from_pretrained passes token=True which requires
+            # a stored HF token even though the repo is public).
+            try:
+                local_path = snapshot_download(
+                    repo_id=CHATTERBOX_TURBO_HF_REPO,
+                    token=None,
+                    allow_patterns=[
+                        "*.safetensors", "*.json", "*.txt", "*.pt", "*.model",
+                    ],
+                )
+            finally:
+                tracker_context.__exit__(None, None, None)
+
+            # Monkey-patch torch.load for CPU loading. The model's .pt files
+            # were saved on CUDA; from_local() doesn't pass map_location
+            # so loading on CPU fails without this.
+            # Load into a local var, apply patches, then publish to
+            # self.model so a failed patch doesn't leave us half-initialised.
+            if device == "cpu":
+                _orig_torch_load = torch.load
+
+                def _patched_load(*args, **kwargs):
+                    kwargs.setdefault("map_location", "cpu")
+                    return _orig_torch_load(*args, **kwargs)
+
+                with ChatterboxTurboTTSBackend._load_lock:
+                    torch.load = _patched_load
+                    try:
+                        model = ChatterboxTurboTTS.from_local(
+                            local_path, device,
+                        )
+                    finally:
+                        torch.load = _orig_torch_load
+            else:
+                model = ChatterboxTurboTTS.from_local(
+                    local_path, device,
+                )
+
+            if not is_cached:
+                progress_manager.mark_complete(model_name)
+                task_manager.complete_download(model_name)
+
+            # Patch float64 → float32 dtype mismatches in upstream chatterbox.
+            # librosa.load returns float64 numpy; multiple upstream code paths
+            # convert it to a torch tensor via torch.from_numpy() without
+            # casting, then matmul it against float32 model weights.
+            # We patch the two known entry points:
+            #
+            # 1. S3Tokenizer.log_mel_spectrogram — the audio tensor from
+            #    librosa hits _mel_filters (float32) in a matmul.
+            # 2. VoiceEncoder.forward — float64 mel spectrograms hit the
+            #    float32 LSTM weights.
+            import types
+
+            # Patch S3Tokenizer (used by s3gen.tokenizer)
+            _tokzr = model.s3gen.tokenizer
+            _orig_log_mel = _tokzr.log_mel_spectrogram.__func__
+
+            def _f32_log_mel(self_tokzr, audio, padding=0):
+                import torch as _torch
+                if _torch.is_tensor(audio):
+                    audio = audio.float()
+                return _orig_log_mel(self_tokzr, audio, padding)
+
+            _tokzr.log_mel_spectrogram = types.MethodType(_f32_log_mel, _tokzr)
+
+            # Patch VoiceEncoder
+            _ve = model.ve
+            _orig_ve_forward = _ve.forward.__func__
+
+            def _f32_ve_forward(self_ve, mels):
+                return _orig_ve_forward(self_ve, mels.float())
+
+            _ve.forward = types.MethodType(_f32_ve_forward, _ve)
+
+            # Only publish after all patches succeed
+            self.model = model
+
+            logger.info("Chatterbox Turbo TTS loaded successfully")
+
+        except ImportError as e:
+            logger.error(
+                "chatterbox-tts package not found. "
+                "Install with: pip install chatterbox-tts"
+            )
+            if not is_cached:
+                progress_manager.mark_error(model_name, str(e))
+                task_manager.error_download(model_name, str(e))
+            raise
+        except Exception as e:
+            logger.error(f"Failed to load Chatterbox Turbo: {e}")
+            if not is_cached:
+                progress_manager.mark_error(model_name, str(e))
+                task_manager.error_download(model_name, str(e))
+            raise
+
+    def unload_model(self) -> None:
+        """Unload model to free memory."""
+        if self.model is not None:
+            device = self._device
+            del self.model
+            self.model = None
+            self._device = None
+            if device == "cuda":
+                import torch
+
+                torch.cuda.empty_cache()
+            logger.info("Chatterbox Turbo unloaded")
+
+    async def create_voice_prompt(
+        self,
+        audio_path: str,
+        reference_text: str,
+        use_cache: bool = True,
+    ) -> Tuple[dict, bool]:
+        """
+        Create voice prompt from reference audio.
+
+        Chatterbox Turbo processes reference audio at generation time, so the
+        prompt just stores the file path.
+        """
+        voice_prompt = {
+            "ref_audio": str(audio_path),
+            "ref_text": reference_text,
+        }
+        return voice_prompt, False
+
+    async def combine_voice_prompts(
+        self,
+        audio_paths: List[str],
+        reference_texts: List[str],
+    ) -> Tuple[np.ndarray, str]:
+        """Combine multiple reference samples."""
+        combined_audio = []
+        for path in audio_paths:
+            audio, _sr = load_audio(path)
+            audio = normalize_audio(audio)
+            combined_audio.append(audio)
+
+        mixed = np.concatenate(combined_audio)
+        mixed = normalize_audio(mixed)
+        combined_text = " ".join(reference_texts)
+        return mixed, combined_text
+
+    async def generate(
+        self,
+        text: str,
+        voice_prompt: dict,
+        language: str = "en",
+        seed: Optional[int] = None,
+        instruct: Optional[str] = None,
+    ) -> Tuple[np.ndarray, int]:
+        """
+        Generate audio using Chatterbox Turbo TTS.
+
+        Supports paralinguistic tags in text: [laugh], [cough], [chuckle], etc.
+
+        Args:
+            text: Text to synthesize (may include paralinguistic tags)
+            voice_prompt: Dict with ref_audio path
+            language: Ignored (Turbo is English-only)
+            seed: Random seed for reproducibility
+            instruct: Unused (protocol compatibility)
+
+        Returns:
+            Tuple of (audio_array, sample_rate)
+        """
+        await self.load_model()
+
+        ref_audio = voice_prompt.get("ref_audio")
+        if ref_audio and not Path(ref_audio).exists():
+            logger.warning(f"Reference audio not found: {ref_audio}")
+            ref_audio = None
+
+        def _generate_sync():
+            import torch
+
+            if seed is not None:
+                torch.manual_seed(seed)
+
+            logger.info("[Chatterbox Turbo] Generating (English)")
+
+            wav = self.model.generate(
+                text,
+                audio_prompt_path=ref_audio,
+                temperature=0.8,
+                top_k=1000,
+                top_p=0.95,
+                repetition_penalty=1.2,
+            )
+
+            # Convert tensor -> numpy
+            if isinstance(wav, torch.Tensor):
+                audio = wav.squeeze().cpu().numpy().astype(np.float32)
+            else:
+                audio = np.asarray(wav, dtype=np.float32)
+
+            sample_rate = (
+                getattr(self.model, "sr", None)
+                or getattr(self.model, "sample_rate", 24000)
+            )
+
+            return audio, sample_rate
+
+        return await asyncio.to_thread(_generate_sync)

backend/backends/mlx_backend.py

@@ -5,8 +5,15 @@ MLX backend implementation for TTS and STT using mlx-audio.
 from typing import Optional, List, Tuple
 import asyncio
 import numpy as np
+import os
 from pathlib import Path
 
+# PATCH: Import and apply offline patch BEFORE any huggingface_hub usage
+# This prevents mlx_audio from making network requests when models are cached
+from ..utils.hf_offline_patch import patch_huggingface_hub_offline, ensure_original_qwen_config_cached
+patch_huggingface_hub_offline()
+ensure_original_qwen_config_cached()
+
 from . import TTSBackend, STTBackend
 from ..utils.cache import get_cache_key, get_cached_voice_prompt, cache_voice_prompt
 from ..utils.audio import normalize_audio, load_audio
@@ -14,6 +21,12 @@ from ..utils.progress import get_progress_manager
 from ..utils.hf_progress import HFProgressTracker, create_hf_progress_callback
 from ..utils.tasks import get_task_manager
 
+LANGUAGE_CODE_TO_NAME = {
+    "zh": "chinese", "en": "english", "ja": "japanese", "ko": "korean",
+    "de": "german", "fr": "french", "ru": "russian", "pt": "portuguese",
+    "es": "spanish", "it": "italian",
+}
+
 
 class MLXTTSBackend:
     """MLX-based TTS backend using mlx-audio."""
@@ -159,15 +172,35 @@ class MLXTTSBackend:
             tracker_context = tracker.patch_download()
             tracker_context.__enter__()
             
+            # PATCH: Force offline mode when model is already cached
+            # This prevents crashes when HuggingFace is unreachable
+            original_hf_hub_offline = os.environ.get("HF_HUB_OFFLINE")
+            if is_cached:
+                os.environ["HF_HUB_OFFLINE"] = "1"
+                print(f"[PATCH] Model {model_size} is cached, forcing HF_HUB_OFFLINE=1 to avoid network requests")
+            
             # Import mlx_audio AFTER patching tqdm
             from mlx_audio.tts import load
             
             # Load MLX model (downloads automatically)
             try:
                 self.model = load(model_path)
+            except Exception as load_error:
+                # If offline mode failed, try with network enabled as fallback
+                if is_cached and "offline" in str(load_error).lower():
+                    print(f"[PATCH] Offline load failed, trying with network: {load_error}")
+                    os.environ.pop("HF_HUB_OFFLINE", None)
+                    self.model = load(model_path)
+                else:
+                    raise
             finally:
                 # Exit the patch context
                 tracker_context.__exit__(None, None, None)
+                # Restore original HF_HUB_OFFLINE setting
+                if original_hf_hub_offline is not None:
+                    os.environ["HF_HUB_OFFLINE"] = original_hf_hub_offline
+                else:
+                    os.environ.pop("HF_HUB_OFFLINE", None)
             
             # Only mark download as complete if we were tracking it
             if not is_cached:
@@ -316,7 +349,8 @@ class MLXTTSBackend:
             # MLX generate() returns a generator yielding GenerationResult objects
             audio_chunks = []
             sample_rate = 24000
-            
+            lang = LANGUAGE_CODE_TO_NAME.get(language, "auto")
+
             # Set seed if provided (MLX uses numpy random)
             if seed is not None:
                 import mlx.core as mx
@@ -344,23 +378,23 @@ class MLXTTSBackend:
                     sig = inspect.signature(self.model.generate)
                     if "ref_audio" in sig.parameters:
                         # Generate with voice cloning
-                        for result in self.model.generate(text, ref_audio=ref_audio, ref_text=ref_text):
+                        for result in self.model.generate(text, ref_audio=ref_audio, ref_text=ref_text, lang_code=lang):
                             audio_chunks.append(np.array(result.audio))
                             sample_rate = result.sample_rate
                     else:
                         # Fallback: generate without voice cloning
-                        for result in self.model.generate(text):
+                        for result in self.model.generate(text, lang_code=lang):
                             audio_chunks.append(np.array(result.audio))
                             sample_rate = result.sample_rate
                 else:
                     # No voice prompt, generate normally
-                    for result in self.model.generate(text):
+                    for result in self.model.generate(text, lang_code=lang):
                         audio_chunks.append(np.array(result.audio))
                         sample_rate = result.sample_rate
             except Exception as e:
                 # If voice cloning fails, try without it
                 print(f"Warning: Voice cloning failed, generating without voice prompt: {e}")
-                for result in self.model.generate(text):
+                for result in self.model.generate(text, lang_code=lang):
                     audio_chunks.append(np.array(result.audio))
                     sample_rate = result.sample_rate
             

backend/backends/pytorch_backend.py

@@ -15,6 +15,12 @@ from ..utils.progress import get_progress_manager
 from ..utils.hf_progress import HFProgressTracker, create_hf_progress_callback
 from ..utils.tasks import get_task_manager
 
+LANGUAGE_CODE_TO_NAME = {
+    "zh": "chinese", "en": "english", "ja": "japanese", "ko": "korean",
+    "de": "german", "fr": "french", "ru": "russian", "pt": "portuguese",
+    "es": "spanish", "it": "italian",
+}
+
 
 class PyTorchTTSBackend:
     """PyTorch-based TTS backend using Qwen3-TTS."""
@@ -359,6 +365,7 @@ class PyTorchTTSBackend:
             wavs, sample_rate = self.model.generate_voice_clone(
                 text=text,
                 voice_clone_prompt=voice_prompt,
+                language=LANGUAGE_CODE_TO_NAME.get(language, "auto"),
                 instruct=instruct,
             )
             return wavs[0], sample_rate
@@ -374,6 +381,7 @@ WHISPER_HF_REPOS = {
     "small": "openai/whisper-small",
     "medium": "openai/whisper-medium",
     "large": "openai/whisper-large-v3",
+    "turbo": "openai/whisper-large-v3-turbo",
 }
 
 
@@ -591,21 +599,20 @@ class PyTorchSTTBackend:
             )
             inputs = inputs.to(self.device)
             
-            # Set language if provided
-            forced_decoder_ids = None
+            # Generate transcription
+            # If language is provided, force it; otherwise let Whisper auto-detect
+            generate_kwargs = {}
             if language:
-                # Support all languages from frontend: en, zh, ja, ko, de, fr, ru, pt, es, it
-                # Whisper supports these and many more
                 forced_decoder_ids = self.processor.get_decoder_prompt_ids(
                     language=language,
                     task="transcribe",
                 )
+                generate_kwargs["forced_decoder_ids"] = forced_decoder_ids
             
-            # Generate transcription
             with torch.no_grad():
                 predicted_ids = self.model.generate(
                     inputs["input_features"],
-                    forced_decoder_ids=forced_decoder_ids,
+                    **generate_kwargs,
                 )
             
             # Decode

backend/main.py

@@ -14,7 +14,6 @@ from datetime import datetime
 import asyncio
 import uvicorn
 import argparse
-import torch
 import tempfile
 import io
 from pathlib import Path
@@ -22,6 +21,18 @@ import uuid
 import asyncio
 import signal
 import os
+
+# Set HSA_OVERRIDE_GFX_VERSION for AMD GPUs that aren't officially listed in ROCm
+# (e.g., RX 6600 is gfx1032 which maps to gfx1030 target)
+# This must be set BEFORE any torch.cuda calls
+if not os.environ.get("HSA_OVERRIDE_GFX_VERSION"):
+    os.environ["HSA_OVERRIDE_GFX_VERSION"] = "10.3.0"
+
+# Suppress noisy MIOpen workspace warnings on AMD GPUs
+if not os.environ.get("MIOPEN_LOG_LEVEL"):
+    os.environ["MIOPEN_LOG_LEVEL"] = "4"
+
+import torch
 from urllib.parse import quote
 
 
@@ -66,10 +77,23 @@ app = FastAPI(
     version=__version__,
 )
 
-# CORS middleware
+# CORS middleware - restrict to known local origins by default.
+# Set VOICEBOX_CORS_ORIGINS env var to a comma-separated list of origins
+# to allow additional origins (e.g. for remote server mode).
+_default_origins = [
+    "http://localhost:5173",     # Vite dev server
+    "http://127.0.0.1:5173",
+    "http://localhost:17493",
+    "http://127.0.0.1:17493",
+    "tauri://localhost",         # Tauri webview (macOS)
+    "https://tauri.localhost",   # Tauri webview (Windows/Linux)
+]
+_env_origins = os.environ.get("VOICEBOX_CORS_ORIGINS", "")
+_cors_origins = _default_origins + [o.strip() for o in _env_origins.split(",") if o.strip()]
+
 app.add_middleware(
     CORSMiddleware,
-    allow_origins=["*"],  # Configure appropriately for production
+    allow_origins=_cors_origins,
     allow_credentials=True,
     allow_methods=["*"],
     allow_headers=["*"],
@@ -222,6 +246,75 @@ async def health():
     )
 
 
+@app.get("/health/filesystem", response_model=models.FilesystemHealthResponse)
+async def filesystem_health():
+    """Check filesystem health: directory existence, write permissions, and disk space."""
+    import shutil
+
+    dirs_to_check = {
+        "generations": config.get_generations_dir(),
+        "profiles": config.get_profiles_dir(),
+        "data": config.get_data_dir(),
+    }
+
+    checks: list[models.DirectoryCheck] = []
+    all_ok = True
+
+    for _label, dir_path in dirs_to_check.items():
+        exists = dir_path.exists()
+        writable = False
+        error = None
+        if exists:
+            # Probe writability with a temp file
+            probe = dir_path / ".voicebox_probe"
+            try:
+                probe.write_text("ok")
+                probe.unlink()
+                writable = True
+            except PermissionError:
+                error = "Permission denied"
+            except OSError as e:
+                error = str(e)
+            finally:
+                try:
+                    probe.unlink(missing_ok=True)
+                except Exception:
+                    pass
+        else:
+            error = "Directory does not exist"
+
+        if not exists or not writable:
+            all_ok = False
+
+        checks.append(
+            models.DirectoryCheck(
+                path=str(dir_path),
+                exists=exists,
+                writable=writable,
+                error=error,
+            )
+        )
+
+    # Disk space for the data directory
+    disk_free_mb = None
+    disk_total_mb = None
+    try:
+        usage = shutil.disk_usage(str(config.get_data_dir()))
+        disk_free_mb = round(usage.free / (1024 * 1024), 1)
+        disk_total_mb = round(usage.total / (1024 * 1024), 1)
+        if disk_free_mb < 500:
+            all_ok = False
+    except OSError:
+        all_ok = False
+
+    return models.FilesystemHealthResponse(
+        healthy=all_ok,
+        disk_free_mb=disk_free_mb,
+        disk_total_mb=disk_total_mb,
+        directories=checks,
+    )
+
+
 # ============================================
 # VOICE PROFILE ENDPOINTS
 # ============================================
@@ -699,6 +792,29 @@ async def generate_speech(
                 )
 
             await tts_model.load_model()
+        elif engine == "chatterbox_turbo":
+            if not tts_model._is_model_cached():
+                model_name = "chatterbox-turbo"
+
+                async def download_chatterbox_turbo_background():
+                    try:
+                        await tts_model.load_model()
+                    except Exception as e:
+                        task_manager.error_download(model_name, str(e))
+
+                task_manager.start_download(model_name)
+                asyncio.create_task(download_chatterbox_turbo_background())
+
+                raise HTTPException(
+                    status_code=202,
+                    detail={
+                        "message": "Chatterbox Turbo model is being downloaded. Please wait and try again.",
+                        "model_name": model_name,
+                        "downloading": True,
+                    },
+                )
+
+            await tts_model.load_model()
 
         # Create voice prompt from profile
         voice_prompt = await profiles.create_voice_prompt_for_profile(
@@ -708,18 +824,29 @@ async def generate_speech(
             engine=engine,
         )
 
-        audio, sample_rate = await tts_model.generate(
+        from .utils.chunked_tts import generate_chunked
+
+        # Resolve per-chunk trim function for engines that need it
+        trim_fn = None
+        if engine in ("chatterbox", "chatterbox_turbo"):
+            from .utils.audio import trim_tts_output
+            trim_fn = trim_tts_output
+
+        audio, sample_rate = await generate_chunked(
+            tts_model,
             data.text,
             voice_prompt,
-            data.language,
-            data.seed,
-            data.instruct,
+            language=data.language,
+            seed=data.seed,
+            instruct=data.instruct,
+            max_chunk_chars=data.max_chunk_chars,
+            crossfade_ms=data.crossfade_ms,
+            trim_fn=trim_fn,
         )
 
-        # Trim trailing silence/hallucination for Chatterbox output
-        if engine == "chatterbox":
-            from .utils.audio import trim_tts_output
-            audio = trim_tts_output(audio, sample_rate)
+        if data.normalize:
+            from .utils.audio import normalize_audio
+            audio = normalize_audio(audio)
 
         # Calculate duration
         duration = len(audio) / sample_rate
@@ -728,7 +855,30 @@ async def generate_speech(
         audio_path = config.get_generations_dir() / f"{generation_id}.wav"
 
         from .utils.audio import save_audio
-        save_audio(audio, str(audio_path), sample_rate)
+        import errno
+
+        try:
+            save_audio(audio, str(audio_path), sample_rate)
+        except BrokenPipeError:
+            raise HTTPException(
+                status_code=500,
+                detail="Audio save failed: broken pipe (the output stream was closed unexpectedly)",
+            )
+        except OSError as save_err:
+            err_no = getattr(save_err, "errno", None) or (
+                getattr(save_err.__cause__, "errno", None)
+                if save_err.__cause__
+                else None
+            )
+            if err_no == errno.ENOENT:
+                msg = f"Audio save failed: directory not found — {audio_path.parent}"
+            elif err_no == errno.EACCES:
+                msg = f"Audio save failed: permission denied — {audio_path.parent}"
+            elif err_no == errno.ENOSPC:
+                msg = "Audio save failed: no disk space remaining"
+            else:
+                msg = f"Audio save failed: {save_err}"
+            raise HTTPException(status_code=500, detail=msg)
 
         # Create history entry
         generation = await history.create_generation(
@@ -798,23 +948,40 @@ async def stream_speech(
                 detail="Chatterbox model is not downloaded yet. Use /generate to trigger a download.",
             )
         await tts_model.load_model()
+    elif engine == "chatterbox_turbo":
+        if not tts_model._is_model_cached():
+            raise HTTPException(
+                status_code=400,
+                detail="Chatterbox Turbo model is not downloaded yet. Use /generate to trigger a download.",
+            )
+        await tts_model.load_model()
 
     voice_prompt = await profiles.create_voice_prompt_for_profile(
         data.profile_id, db, engine=engine,
     )
 
-    audio, sample_rate = await tts_model.generate(
+    from .utils.chunked_tts import generate_chunked
+
+    trim_fn = None
+    if engine in ("chatterbox", "chatterbox_turbo"):
+        from .utils.audio import trim_tts_output
+        trim_fn = trim_tts_output
+
+    audio, sample_rate = await generate_chunked(
+        tts_model,
         data.text,
         voice_prompt,
-        data.language,
-        data.seed,
-        data.instruct,
+        language=data.language,
+        seed=data.seed,
+        instruct=data.instruct,
+        max_chunk_chars=data.max_chunk_chars,
+        crossfade_ms=data.crossfade_ms,
+        trim_fn=trim_fn,
     )
 
-    # Trim trailing silence/hallucination for Chatterbox output
-    if engine == "chatterbox":
-        from .utils.audio import trim_tts_output
-        audio = trim_tts_output(audio, sample_rate)
+    if data.normalize:
+        from .utils.audio import normalize_audio
+        audio = normalize_audio(audio)
 
     wav_bytes = tts.audio_to_wav_bytes(audio, sample_rate)
 
@@ -1022,11 +1189,12 @@ async def transcribe_audio(
         # Transcribe
         whisper_model = transcribe.get_whisper_model()
 
-        # Check if Whisper model is downloaded (uses default size "base")
+        # Check if Whisper model is downloaded
         model_size = whisper_model.model_size
-        # Map model sizes to HF repo IDs (whisper-large needs -v3 suffix)
+        # Map model sizes to HF repo IDs (some need special suffixes)
         whisper_hf_repos = {
             "large": "openai/whisper-large-v3",
+            "turbo": "openai/whisper-large-v3-turbo",
         }
         model_name = whisper_hf_repos.get(model_size, f"openai/whisper-{model_size}")
 
@@ -1332,7 +1500,7 @@ async def load_model(model_size: str = "1.7B"):
 
 @app.post("/models/unload")
 async def unload_model():
-    """Unload TTS model to free memory."""
+    """Unload the default Qwen TTS model to free memory."""
     try:
         tts.unload_tts_model()
         return {"message": "Model unloaded successfully"}
@@ -1340,6 +1508,71 @@ async def unload_model():
         raise HTTPException(status_code=500, detail=str(e))
 
 
+@app.post("/models/{model_name}/unload")
+async def unload_model_by_name(model_name: str):
+    """Unload a specific model from memory without deleting it from disk."""
+    # Map of model_name -> (model_type, model_size)
+    model_types = {
+        "qwen-tts-1.7B": ("tts", "1.7B"),
+        "qwen-tts-0.6B": ("tts", "0.6B"),
+        "luxtts": ("luxtts", "default"),
+        "chatterbox-tts": ("chatterbox", "default"),
+        "chatterbox-turbo": ("chatterbox_turbo", "default"),
+        "whisper-base": ("whisper", "base"),
+        "whisper-small": ("whisper", "small"),
+        "whisper-medium": ("whisper", "medium"),
+        "whisper-large": ("whisper", "large"),
+        "whisper-turbo": ("whisper", "turbo"),
+    }
+
+    if model_name not in model_types:
+        raise HTTPException(status_code=400, detail=f"Unknown model: {model_name}")
+
+    model_type, model_size = model_types[model_name]
+
+    try:
+        if model_type == "tts":
+            tts_model = tts.get_tts_model()
+            loaded_size = getattr(
+                tts_model, "_current_model_size", None
+            ) or getattr(tts_model, "model_size", None)
+            if tts_model.is_loaded() and loaded_size == model_size:
+                tts.unload_tts_model()
+            else:
+                return {"message": f"Model {model_name} is not loaded"}
+        elif model_type == "luxtts":
+            from .backends import get_tts_backend_for_engine
+            backend = get_tts_backend_for_engine("luxtts")
+            if backend.is_loaded():
+                backend.unload_model()
+            else:
+                return {"message": f"Model {model_name} is not loaded"}
+        elif model_type == "chatterbox":
+            from .backends import get_tts_backend_for_engine
+            backend = get_tts_backend_for_engine("chatterbox")
+            if backend.is_loaded():
+                backend.unload_model()
+            else:
+                return {"message": f"Model {model_name} is not loaded"}
+        elif model_type == "chatterbox_turbo":
+            from .backends import get_tts_backend_for_engine
+            backend = get_tts_backend_for_engine("chatterbox_turbo")
+            if backend.is_loaded():
+                backend.unload_model()
+            else:
+                return {"message": f"Model {model_name} is not loaded"}
+        elif model_type == "whisper":
+            whisper_model = transcribe.get_whisper_model()
+            if whisper_model.is_loaded() and whisper_model.model_size == model_size:
+                transcribe.unload_whisper_model()
+            else:
+                return {"message": f"Model {model_name} is not loaded"}
+
+        return {"message": f"Model {model_name} unloaded successfully"}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e)) from e
+
+
 @app.get("/models/progress/{model_name}")
 async def get_model_progress(model_name: str):
     """Get model download progress via Server-Sent Events."""
@@ -1363,6 +1596,141 @@ async def get_model_progress(model_name: str):
     )
 
 
+@app.get("/models/cache-dir")
+async def get_models_cache_dir():
+    """Get the path to the HuggingFace model cache directory."""
+    from huggingface_hub import constants as hf_constants
+    return {"path": str(Path(hf_constants.HF_HUB_CACHE))}
+
+
+def _get_dir_size(path: Path) -> int:
+    """Get total size of a directory in bytes."""
+    total = 0
+    for f in path.rglob("*"):
+        if f.is_file():
+            total += f.stat().st_size
+    return total
+
+
+def _copy_with_progress(src: Path, dst: Path, progress_manager, copied_so_far: int, total_bytes: int) -> int:
+    """Copy a directory tree with byte-level progress tracking."""
+    import shutil
+    dst.mkdir(parents=True, exist_ok=True)
+    for item in src.iterdir():
+        dest_item = dst / item.name
+        if item.is_dir():
+            copied_so_far = _copy_with_progress(item, dest_item, progress_manager, copied_so_far, total_bytes)
+        else:
+            size = item.stat().st_size
+            shutil.copy2(str(item), str(dest_item))
+            copied_so_far += size
+            progress_manager.update_progress(
+                "migration", copied_so_far, total_bytes,
+                filename=item.name, status="downloading",
+            )
+    return copied_so_far
+
+
+@app.post("/models/migrate")
+async def migrate_models(request: models.ModelMigrateRequest):
+    """Move all downloaded models to a new directory with byte-level progress via SSE."""
+    import shutil
+    from huggingface_hub import constants as hf_constants
+
+    source = Path(hf_constants.HF_HUB_CACHE)
+    destination = Path(request.destination)
+
+    if not source.exists():
+        raise HTTPException(status_code=404, detail="Current model cache directory not found")
+
+    model_dirs = [d for d in source.iterdir() if d.name.startswith("models--") and d.is_dir()]
+    if not model_dirs:
+        return {"moved": 0, "errors": [], "source": str(source), "destination": str(destination)}
+
+    destination.mkdir(parents=True, exist_ok=True)
+
+    progress_manager = get_progress_manager()
+
+    # Check if source and destination are on the same filesystem (rename is instant)
+    same_fs = False
+    try:
+        same_fs = source.stat().st_dev == destination.stat().st_dev
+    except OSError:
+        pass
+
+    async def migrate_background():
+        moved = 0
+        errors = []
+        try:
+            if same_fs:
+                # Same filesystem: rename is instant, just track model count
+                total = len(model_dirs)
+                for i, item in enumerate(model_dirs):
+                    dest_item = destination / item.name
+                    try:
+                        if dest_item.exists():
+                            shutil.rmtree(dest_item)
+                        shutil.move(str(item), str(dest_item))
+                        moved += 1
+                        progress_manager.update_progress(
+                            "migration", i + 1, total,
+                            filename=item.name, status="downloading",
+                        )
+                    except Exception as e:
+                        errors.append(f"{item.name}: {str(e)}")
+            else:
+                # Cross-filesystem: copy with byte-level progress, then delete source
+                total_bytes = sum(_get_dir_size(d) for d in model_dirs)
+                progress_manager.update_progress("migration", 0, total_bytes, filename="Calculating...", status="downloading")
+
+                copied = 0
+                for item in model_dirs:
+                    dest_item = destination / item.name
+                    try:
+                        if dest_item.exists():
+                            shutil.rmtree(dest_item)
+                        copied = await asyncio.to_thread(
+                            _copy_with_progress, item, dest_item, progress_manager, copied, total_bytes
+                        )
+                        # Remove source after successful copy
+                        await asyncio.to_thread(shutil.rmtree, str(item))
+                        moved += 1
+                    except Exception as e:
+                        errors.append(f"{item.name}: {str(e)}")
+
+            progress_manager.update_progress("migration", 1, 1, status="complete")
+            progress_manager.mark_complete("migration")
+        except Exception as e:
+            progress_manager.update_progress("migration", 0, 0, status="error")
+            progress_manager.mark_error("migration", str(e))
+
+    _create_background_task(migrate_background())
+
+    return {"source": str(source), "destination": str(destination)}
+
+
+@app.get("/models/migrate/progress")
+async def get_migration_progress():
+    """Get model migration progress via Server-Sent Events."""
+    from fastapi.responses import StreamingResponse
+
+    progress_manager = get_progress_manager()
+
+    async def event_generator():
+        async for event in progress_manager.subscribe("migration"):
+            yield event
+
+    return StreamingResponse(
+        event_generator(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache",
+            "Connection": "keep-alive",
+            "X-Accel-Buffering": "no",
+        },
+    )
+
+
 @app.get("/models/status", response_model=models.ModelStatusListResponse)
 async def get_model_status():
     """Get status of all available models."""
@@ -1386,7 +1754,10 @@ async def get_model_status():
         """Check if TTS model is loaded with specific size."""
         try:
             tts_model = tts.get_tts_model()
-            return tts_model.is_loaded() and getattr(tts_model, 'model_size', None) == model_size
+            loaded_size = getattr(
+                tts_model, "_current_model_size", None
+            ) or getattr(tts_model, "model_size", None)
+            return tts_model.is_loaded() and loaded_size == model_size
         except Exception:
             return False
     
@@ -1433,6 +1804,15 @@ async def get_model_status():
         except Exception:
             return False
 
+    # Check if Chatterbox Turbo backend is loaded
+    def check_chatterbox_turbo_loaded():
+        try:
+            from .backends import get_tts_backend_for_engine
+            backend = get_tts_backend_for_engine("chatterbox_turbo")
+            return backend.is_loaded()
+        except Exception:
+            return False
+
     model_configs = [
         {
             "model_name": "qwen-tts-1.7B",
@@ -1462,6 +1842,13 @@ async def get_model_status():
             "model_size": "default",
             "check_loaded": check_chatterbox_loaded,
         },
+        {
+            "model_name": "chatterbox-turbo",
+            "display_name": "Chatterbox Turbo (English, Tags)",
+            "hf_repo_id": "ResembleAI/chatterbox-turbo",
+            "model_size": "default",
+            "check_loaded": check_chatterbox_turbo_loaded,
+        },
         {
             "model_name": "whisper-base",
             "display_name": "Whisper Base",
@@ -1490,6 +1877,13 @@ async def get_model_status():
             "model_size": "large",
             "check_loaded": lambda: check_whisper_loaded("large"),
         },
+        {
+            "model_name": "whisper-turbo",
+            "display_name": "Whisper Turbo",
+            "hf_repo_id": "openai/whisper-large-v3-turbo",
+            "model_size": "turbo",
+            "check_loaded": lambda: check_whisper_loaded("turbo"),
+        },
     ]
     
     # Build a mapping of model_name -> hf_repo_id so we can check if shared repos are downloading
@@ -1668,6 +2062,10 @@ async def trigger_model_download(request: models.ModelDownloadRequest):
             "model_size": "default",
             "load_func": lambda: get_tts_backend_for_engine("chatterbox").load_model(),
         },
+        "chatterbox-turbo": {
+            "model_size": "default",
+            "load_func": lambda: get_tts_backend_for_engine("chatterbox_turbo").load_model(),
+        },
         "whisper-base": {
             "model_size": "base",
             "load_func": lambda: transcribe.get_whisper_model().load_model("base"),
@@ -1684,6 +2082,10 @@ async def trigger_model_download(request: models.ModelDownloadRequest):
             "model_size": "large",
             "load_func": lambda: transcribe.get_whisper_model().load_model("large"),
         },
+        "whisper-turbo": {
+            "model_size": "turbo",
+            "load_func": lambda: transcribe.get_whisper_model().load_model("turbo"),
+        },
     }
     
     if request.model_name not in model_configs:
@@ -1790,6 +2192,11 @@ async def delete_model(model_name: str):
             "model_size": "default",
             "model_type": "chatterbox",
         },
+        "chatterbox-turbo": {
+            "hf_repo_id": "ResembleAI/chatterbox-turbo",
+            "model_size": "default",
+            "model_type": "chatterbox_turbo",
+        },
         "whisper-base": {
             "hf_repo_id": "openai/whisper-base",
             "model_size": "base",
@@ -1810,6 +2217,11 @@ async def delete_model(model_name: str):
             "model_size": "large",
             "model_type": "whisper",
         },
+        "whisper-turbo": {
+            "hf_repo_id": "openai/whisper-large-v3-turbo",
+            "model_size": "turbo",
+            "model_type": "whisper",
+        },
     }
 
     if model_name not in model_configs:
@@ -1822,7 +2234,10 @@ async def delete_model(model_name: str):
         # Check if model is loaded and unload it first
         if config["model_type"] == "tts":
             tts_model = tts.get_tts_model()
-            if tts_model.is_loaded() and tts_model.model_size == config["model_size"]:
+            loaded_size = getattr(
+                tts_model, "_current_model_size", None
+            ) or getattr(tts_model, "model_size", None)
+            if tts_model.is_loaded() and loaded_size == config["model_size"]:
                 tts.unload_tts_model()
         elif config["model_type"] == "luxtts":
             from .backends import get_tts_backend_for_engine
@@ -1834,6 +2249,11 @@ async def delete_model(model_name: str):
             chatterbox = get_tts_backend_for_engine("chatterbox")
             if chatterbox.is_loaded():
                 chatterbox.unload_model()
+        elif config["model_type"] == "chatterbox_turbo":
+            from .backends import get_tts_backend_for_engine
+            turbo = get_tts_backend_for_engine("chatterbox_turbo")
+            if turbo.is_loaded():
+                turbo.unload_model()
         elif config["model_type"] == "whisper":
             whisper_model = transcribe.get_whisper_model()
             if whisper_model.is_loaded() and whisper_model.model_size == config["model_size"]:
@@ -2044,7 +2464,12 @@ def _get_gpu_status() -> str:
     """Get GPU availability status."""
     backend_type = get_backend_type()
     if torch.cuda.is_available():
-        return f"CUDA ({torch.cuda.get_device_name(0)})"
+        device_name = torch.cuda.get_device_name(0)
+        # Check if this is ROCm (AMD) or CUDA (NVIDIA)
+        is_rocm = hasattr(torch.version, 'hip') and torch.version.hip is not None
+        if is_rocm:
+            return f"ROCm ({device_name})"
+        return f"CUDA ({device_name})"
     elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
         return "MPS (Apple Silicon)"
     elif backend_type == "mlx":

backend/models.py

@@ -11,7 +11,7 @@ class VoiceProfileCreate(BaseModel):
     """Request model for creating a voice profile."""
     name: str = Field(..., min_length=1, max_length=100)
     description: Optional[str] = Field(None, max_length=500)
-    language: str = Field(default="en", pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he)$")
+    language: str = Field(default="en", pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he|ar|da|el|fi|hi|ms|nl|no|pl|sv|sw|tr)$")
 
 
 class VoiceProfileResponse(BaseModel):
@@ -52,12 +52,15 @@ class ProfileSampleResponse(BaseModel):
 class GenerationRequest(BaseModel):
     """Request model for voice generation."""
     profile_id: str
-    text: str = Field(..., min_length=1, max_length=5000)
+    text: str = Field(..., min_length=1, max_length=50000)
     language: str = Field(default="en", pattern="^(zh|en|ja|ko|de|fr|ru|pt|es|it|he)$")
     seed: Optional[int] = Field(None, ge=0)
     model_size: Optional[str] = Field(default="1.7B", pattern="^(1\\.7B|0\\.6B)$")
     instruct: Optional[str] = Field(None, max_length=500)
-    engine: Optional[str] = Field(default="qwen", pattern="^(qwen|luxtts|chatterbox)$")
+    engine: Optional[str] = Field(default="qwen", pattern="^(qwen|luxtts|chatterbox|chatterbox_turbo)$")
+    max_chunk_chars: int = Field(default=800, ge=100, le=5000, description="Max characters per chunk for long text splitting")
+    crossfade_ms: int = Field(default=50, ge=0, le=500, description="Crossfade duration in ms between chunks (0 for hard cut)")
+    normalize: bool = Field(default=True, description="Normalize output audio volume")
 
 
 class GenerationResponse(BaseModel):
@@ -131,6 +134,22 @@ class HealthResponse(BaseModel):
     backend_variant: Optional[str] = None  # Binary variant (cpu or cuda)
 
 
+class DirectoryCheck(BaseModel):
+    """Health status for a single directory."""
+    path: str
+    exists: bool
+    writable: bool
+    error: Optional[str] = None
+
+
+class FilesystemHealthResponse(BaseModel):
+    """Response model for filesystem health check."""
+    healthy: bool
+    disk_free_mb: Optional[float] = None
+    disk_total_mb: Optional[float] = None
+    directories: List[DirectoryCheck]
+
+
 class ModelStatus(BaseModel):
     """Response model for model status."""
     model_name: str
@@ -152,6 +171,11 @@ class ModelDownloadRequest(BaseModel):
     model_name: str
 
 
+class ModelMigrateRequest(BaseModel):
+    """Request model for migrating models to a new directory."""
+    destination: str
+
+
 class ActiveDownloadTask(BaseModel):
     """Response model for active download task."""
     model_name: str

backend/tests/test_cors.py

@@ -0,0 +1,162 @@
+"""
+Tests for CORS origin restrictions.
+
+Validates that the CORS middleware only allows known local origins
+and respects the VOICEBOX_CORS_ORIGINS environment variable.
+
+Uses a minimal FastAPI app that mirrors the exact CORS configuration
+from backend/main.py, so tests run without heavy ML dependencies.
+
+Usage:
+    pip install httpx pytest fastapi starlette
+    python -m pytest backend/tests/test_cors.py -v
+"""
+
+import os
+import pytest
+from unittest.mock import patch
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+from starlette.testclient import TestClient
+
+
+def _build_app(env_origins: str = "") -> FastAPI:
+    """
+    Build a minimal FastAPI app with the same CORS logic as backend/main.py.
+
+    This mirrors the exact code in main.py so the test validates the real
+    configuration without needing torch/numpy/transformers installed.
+    """
+    app = FastAPI()
+
+    _default_origins = [
+        "http://localhost:5173",
+        "http://127.0.0.1:5173",
+        "http://localhost:17493",
+        "http://127.0.0.1:17493",
+        "tauri://localhost",
+        "https://tauri.localhost",
+    ]
+    _cors_origins = _default_origins + [o.strip() for o in env_origins.split(",") if o.strip()]
+
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=_cors_origins,
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+
+    @app.get("/health")
+    async def health():
+        return {"status": "ok"}
+
+    return app
+
+
+@pytest.fixture()
+def client():
+    return TestClient(_build_app())
+
+
+@pytest.fixture()
+def client_with_custom_origins():
+    return TestClient(_build_app("https://custom.example.com,https://other.example.com"))
+
+
+def _get_with_origin(client: TestClient, origin: str) -> dict:
+    """Send a GET with Origin header, return response headers."""
+    response = client.get("/health", headers={"Origin": origin})
+    return dict(response.headers)
+
+
+def _preflight(client: TestClient, origin: str) -> dict:
+    """Send CORS preflight OPTIONS request, return response headers."""
+    response = client.options(
+        "/health",
+        headers={
+            "Origin": origin,
+            "Access-Control-Request-Method": "GET",
+        },
+    )
+    return dict(response.headers)
+
+
+class TestCORSDefaultOrigins:
+    """CORS should allow known local origins and block everything else."""
+
+    @pytest.mark.parametrize("origin", [
+        "http://localhost:5173",
+        "http://127.0.0.1:5173",
+        "http://localhost:17493",
+        "http://127.0.0.1:17493",
+        "tauri://localhost",
+        "https://tauri.localhost",
+    ])
+    def test_allowed_origins(self, client, origin):
+        headers = _get_with_origin(client, origin)
+        assert headers.get("access-control-allow-origin") == origin
+
+    @pytest.mark.parametrize("origin", [
+        "http://evil.com",
+        "http://localhost:9999",
+        "https://attacker.example.com",
+        "null",
+    ])
+    def test_blocked_origins(self, client, origin):
+        headers = _get_with_origin(client, origin)
+        assert "access-control-allow-origin" not in headers
+
+    def test_preflight_allowed(self, client):
+        headers = _preflight(client, "http://localhost:5173")
+        assert headers.get("access-control-allow-origin") == "http://localhost:5173"
+
+    def test_preflight_blocked(self, client):
+        headers = _preflight(client, "http://evil.com")
+        assert "access-control-allow-origin" not in headers
+
+    def test_credentials_header_present(self, client):
+        headers = _get_with_origin(client, "http://localhost:5173")
+        assert headers.get("access-control-allow-credentials") == "true"
+
+
+class TestCORSCustomOrigins:
+    """VOICEBOX_CORS_ORIGINS env var should extend the allowlist."""
+
+    def test_custom_origin_allowed(self, client_with_custom_origins):
+        headers = _get_with_origin(client_with_custom_origins, "https://custom.example.com")
+        assert headers.get("access-control-allow-origin") == "https://custom.example.com"
+
+    def test_other_custom_origin_allowed(self, client_with_custom_origins):
+        headers = _get_with_origin(client_with_custom_origins, "https://other.example.com")
+        assert headers.get("access-control-allow-origin") == "https://other.example.com"
+
+    def test_default_origins_still_work(self, client_with_custom_origins):
+        headers = _get_with_origin(client_with_custom_origins, "http://localhost:5173")
+        assert headers.get("access-control-allow-origin") == "http://localhost:5173"
+
+    def test_unlisted_origin_still_blocked(self, client_with_custom_origins):
+        headers = _get_with_origin(client_with_custom_origins, "http://evil.com")
+        assert "access-control-allow-origin" not in headers
+
+
+class TestCORSEnvVarParsing:
+    """Edge cases for VOICEBOX_CORS_ORIGINS parsing."""
+
+    def test_empty_env_var(self):
+        app = _build_app("")
+        client = TestClient(app)
+        headers = _get_with_origin(client, "http://evil.com")
+        assert "access-control-allow-origin" not in headers
+
+    def test_whitespace_trimmed(self):
+        app = _build_app("  https://spaced.example.com  ")
+        client = TestClient(app)
+        headers = _get_with_origin(client, "https://spaced.example.com")
+        assert headers.get("access-control-allow-origin") == "https://spaced.example.com"
+
+    def test_trailing_comma_ignored(self):
+        app = _build_app("https://one.example.com,")
+        client = TestClient(app)
+        headers = _get_with_origin(client, "https://one.example.com")
+        assert headers.get("access-control-allow-origin") == "https://one.example.com"

backend/utils/audio.py

@@ -70,14 +70,44 @@ def save_audio(
     sample_rate: int = 24000,
 ) -> None:
     """
-    Save audio file.
-    
+    Save audio file with atomic write and error handling.
+
+    Writes to a temporary file first, then atomically renames to the
+    target path.  This prevents corrupted/partial WAV files if the
+    process is interrupted mid-write.
+
     Args:
         audio: Audio array
         path: Output path
         sample_rate: Sample rate
+
+    Raises:
+        OSError: If file cannot be written
     """
-    sf.write(path, audio, sample_rate)
+    from pathlib import Path
+    import os
+
+    temp_path = f"{path}.tmp"
+    try:
+        # Ensure parent directory exists
+        Path(path).parent.mkdir(parents=True, exist_ok=True)
+
+        # Write to temporary file first (explicit format since .tmp
+        # extension is not recognised by soundfile)
+        sf.write(temp_path, audio, sample_rate, format='WAV')
+
+        # Atomic rename to final path
+        os.replace(temp_path, path)
+
+    except Exception as e:
+        # Clean up temp file on failure
+        try:
+            if Path(temp_path).exists():
+                Path(temp_path).unlink()
+        except Exception:
+            pass  # Best effort cleanup
+
+        raise OSError(f"Failed to save audio to {path}: {e}") from e
 
 
 def trim_tts_output(

backend/utils/chunked_tts.py

@@ -0,0 +1,302 @@
+"""
+Chunked TTS generation utilities.
+
+Splits long text into sentence-boundary chunks, generates audio per-chunk
+via any TTSBackend, and concatenates with crossfade.  All logic is
+engine-agnostic — it wraps the standard ``TTSBackend.generate()`` interface.
+
+Short text (≤ max_chunk_chars) uses the single-shot fast path with zero
+overhead.
+"""
+
+import logging
+import re
+from typing import List, Tuple
+
+import numpy as np
+
+logger = logging.getLogger("voicebox.chunked-tts")
+
+# Default chunk size in characters.  Can be overridden per-request via
+# the ``max_chunk_chars`` field on GenerationRequest.
+DEFAULT_MAX_CHUNK_CHARS = 800
+
+# Common abbreviations that should NOT be treated as sentence endings.
+# Lowercase for case-insensitive matching.
+_ABBREVIATIONS = frozenset(
+    {
+        "mr",
+        "mrs",
+        "ms",
+        "dr",
+        "prof",
+        "sr",
+        "jr",
+        "st",
+        "ave",
+        "blvd",
+        "inc",
+        "ltd",
+        "corp",
+        "dept",
+        "est",
+        "approx",
+        "vs",
+        "etc",
+        "e.g",
+        "i.e",
+        "a.m",
+        "p.m",
+        "u.s",
+        "u.s.a",
+        "u.k",
+    }
+)
+
+# Paralinguistic tags used by Chatterbox Turbo.  The splitter must never
+# cut inside one of these.
+_PARA_TAG_RE = re.compile(r"\[[^\]]*\]")
+
+
+# ---------------------------------------------------------------------------
+# Text splitting
+# ---------------------------------------------------------------------------
+
+
+def split_text_into_chunks(text: str, max_chars: int = DEFAULT_MAX_CHUNK_CHARS) -> List[str]:
+    """Split *text* at natural boundaries into chunks of at most *max_chars*.
+
+    Priority: sentence-end (``.!?`` not preceded by an abbreviation and not
+    inside brackets) → clause boundary (``;:,—``) → whitespace → hard cut.
+
+    Paralinguistic tags like ``[laugh]`` are treated as atomic and will not
+    be split across chunks.
+    """
+    text = text.strip()
+    if not text:
+        return []
+    if len(text) <= max_chars:
+        return [text]
+
+    chunks: List[str] = []
+    remaining = text
+
+    while remaining:
+        remaining = remaining.lstrip()
+        if not remaining:
+            break
+        if len(remaining) <= max_chars:
+            chunks.append(remaining)
+            break
+
+        segment = remaining[:max_chars]
+
+        # Try to split at the last real sentence ending
+        split_pos = _find_last_sentence_end(segment)
+        if split_pos == -1:
+            split_pos = _find_last_clause_boundary(segment)
+        if split_pos == -1:
+            split_pos = segment.rfind(" ")
+        if split_pos == -1:
+            # Absolute fallback: hard cut but avoid splitting inside a tag
+            split_pos = _safe_hard_cut(segment, max_chars)
+
+        chunk = remaining[: split_pos + 1].strip()
+        if chunk:
+            chunks.append(chunk)
+        remaining = remaining[split_pos + 1 :]
+
+    return chunks
+
+
+def _find_last_sentence_end(text: str) -> int:
+    """Return the index of the last sentence-ending punctuation in *text*.
+
+    Skips periods that follow common abbreviations (``Dr.``, ``Mr.``, etc.)
+    and periods inside bracket tags (``[laugh]``).  Also handles CJK
+    sentence-ending punctuation (``。！？``).
+    """
+    best = -1
+    # ASCII sentence ends
+    for m in re.finditer(r"[.!?](?:\s|$)", text):
+        pos = m.start()
+        char = text[pos]
+        # Skip periods after abbreviations
+        if char == ".":
+            # Walk backwards to find the preceding word
+            word_start = pos - 1
+            while word_start >= 0 and text[word_start].isalpha():
+                word_start -= 1
+            word = text[word_start + 1 : pos].lower()
+            if word in _ABBREVIATIONS:
+                continue
+            # Skip decimal numbers (digit immediately before the period)
+            if word_start >= 0 and text[word_start].isdigit():
+                continue
+        # Skip if we're inside a bracket tag
+        if _inside_bracket_tag(text, pos):
+            continue
+        best = pos
+    # CJK sentence-ending punctuation
+    for m in re.finditer(r"[\u3002\uff01\uff1f]", text):
+        if m.start() > best:
+            best = m.start()
+    return best
+
+
+def _find_last_clause_boundary(text: str) -> int:
+    """Return the index of the last clause-boundary punctuation."""
+    best = -1
+    for m in re.finditer(r"[;:,\u2014](?:\s|$)", text):
+        pos = m.start()
+        # Skip if inside a bracket tag
+        if _inside_bracket_tag(text, pos):
+            continue
+        best = pos
+    return best
+
+
+def _inside_bracket_tag(text: str, pos: int) -> bool:
+    """Return True if *pos* falls inside a ``[...]`` tag."""
+    for m in _PARA_TAG_RE.finditer(text):
+        if m.start() < pos < m.end():
+            return True
+    return False
+
+
+def _safe_hard_cut(segment: str, max_chars: int) -> int:
+    """Find a hard-cut position that doesn't split a ``[tag]``."""
+    cut = max_chars - 1
+    # Check if the cut falls inside a bracket tag; if so, move before it
+    for m in _PARA_TAG_RE.finditer(segment):
+        if m.start() < cut < m.end():
+            return m.start() - 1 if m.start() > 0 else cut
+    return cut
+
+
+# ---------------------------------------------------------------------------
+# Audio concatenation
+# ---------------------------------------------------------------------------
+
+
+def concatenate_audio_chunks(
+    chunks: List[np.ndarray],
+    sample_rate: int,
+    crossfade_ms: int = 50,
+) -> np.ndarray:
+    """Concatenate audio arrays with a short crossfade to eliminate clicks.
+
+    Each chunk is expected to be a 1-D float32 ndarray at *sample_rate* Hz.
+    """
+    if not chunks:
+        return np.array([], dtype=np.float32)
+    if len(chunks) == 1:
+        return chunks[0]
+
+    crossfade_samples = int(sample_rate * crossfade_ms / 1000)
+    result = np.array(chunks[0], dtype=np.float32, copy=True)
+
+    for chunk in chunks[1:]:
+        if len(chunk) == 0:
+            continue
+        overlap = min(crossfade_samples, len(result), len(chunk))
+        if overlap > 0:
+            fade_out = np.linspace(1.0, 0.0, overlap, dtype=np.float32)
+            fade_in = np.linspace(0.0, 1.0, overlap, dtype=np.float32)
+            result[-overlap:] = result[-overlap:] * fade_out + chunk[:overlap] * fade_in
+            result = np.concatenate([result, chunk[overlap:]])
+        else:
+            result = np.concatenate([result, chunk])
+
+    return result
+
+
+# ---------------------------------------------------------------------------
+# Engine-agnostic chunked generation
+# ---------------------------------------------------------------------------
+
+
+async def generate_chunked(
+    backend,
+    text: str,
+    voice_prompt: dict,
+    language: str = "en",
+    seed: int | None = None,
+    instruct: str | None = None,
+    max_chunk_chars: int = DEFAULT_MAX_CHUNK_CHARS,
+    crossfade_ms: int = 50,
+    trim_fn=None,
+) -> Tuple[np.ndarray, int]:
+    """Generate audio with automatic chunking for long text.
+
+    For text shorter than *max_chunk_chars* this is a thin wrapper around
+    ``backend.generate()`` with zero overhead.
+
+    For longer text the input is split at natural sentence boundaries,
+    each chunk is generated independently, optionally trimmed (useful for
+    Chatterbox engines that hallucinate trailing noise), and the results
+    are concatenated with a crossfade (or hard cut if *crossfade_ms* is 0).
+
+    Parameters
+    ----------
+    backend : TTSBackend
+        Any backend implementing the ``generate()`` protocol.
+    text : str
+        Input text (may be arbitrarily long).
+    voice_prompt, language, seed, instruct
+        Forwarded to ``backend.generate()`` verbatim.
+    max_chunk_chars : int
+        Maximum characters per chunk (default 800).
+    crossfade_ms : int
+        Crossfade duration in milliseconds between chunks.  0 for a hard
+        cut with no overlap (default 50).
+    trim_fn : callable | None
+        Optional ``(audio, sample_rate) -> audio`` post-processing
+        function applied to each chunk before concatenation (e.g.
+        ``trim_tts_output`` for Chatterbox engines).
+
+    Returns
+    -------
+    (audio, sample_rate) : Tuple[np.ndarray, int]
+    """
+    chunks = split_text_into_chunks(text, max_chunk_chars)
+
+    if len(chunks) <= 1:
+        # Short text — single-shot fast path
+        audio, sample_rate = await backend.generate(
+            text, voice_prompt, language, seed, instruct,
+        )
+        if trim_fn is not None:
+            audio = trim_fn(audio, sample_rate)
+        return audio, sample_rate
+
+    # Long text — chunked generation
+    logger.info(
+        "Splitting %d chars into %d chunks (max %d chars each)",
+        len(text), len(chunks), max_chunk_chars,
+    )
+    audio_chunks: List[np.ndarray] = []
+    sample_rate: int | None = None
+
+    for i, chunk_text in enumerate(chunks):
+        logger.info(
+            "Generating chunk %d/%d (%d chars)",
+            i + 1, len(chunks), len(chunk_text),
+        )
+        # Vary the seed per chunk to avoid correlated RNG artefacts,
+        # but keep it deterministic so the same (text, seed) pair
+        # always produces the same output.
+        chunk_seed = (seed + i) if seed is not None else None
+
+        chunk_audio, chunk_sr = await backend.generate(
+            chunk_text, voice_prompt, language, chunk_seed, instruct,
+        )
+        if trim_fn is not None:
+            chunk_audio = trim_fn(chunk_audio, chunk_sr)
+
+        audio_chunks.append(np.asarray(chunk_audio, dtype=np.float32))
+        if sample_rate is None:
+            sample_rate = chunk_sr
+
+    audio = concatenate_audio_chunks(audio_chunks, sample_rate, crossfade_ms=crossfade_ms)
+    return audio, sample_rate

backend/utils/hf_offline_patch.py

@@ -0,0 +1,100 @@
+"""
+Monkey patch for huggingface_hub to force offline mode with cached models.
+This prevents mlx_audio from making network requests when models are already downloaded.
+"""
+
+import os
+from pathlib import Path
+from typing import Optional, Union
+
+
+def patch_huggingface_hub_offline():
+    """
+    Monkey-patch huggingface_hub to force offline mode.
+    This must be called BEFORE importing mlx_audio.
+    """
+    try:
+        import huggingface_hub
+        from huggingface_hub import constants as hf_constants
+        from huggingface_hub.file_download import _try_to_load_from_cache
+        
+        # Store original function
+        original_try_load = _try_to_load_from_cache
+        
+        def _patched_try_to_load_from_cache(
+            repo_id: str,
+            filename: str,
+            cache_dir: Union[str, Path, None] = None,
+            revision: Optional[str] = None,
+            repo_type: Optional[str] = None,
+        ):
+            """
+            Patched version that forces offline mode.
+            Returns None if not cached (instead of making network request).
+            """
+            # Always use the original function, but we're already in HF_HUB_OFFLINE mode
+            result = original_try_load(
+                repo_id=repo_id,
+                filename=filename,
+                cache_dir=cache_dir,
+                revision=revision,
+                repo_type=repo_type,
+            )
+            
+            if result is None:
+                # File not in cache - log this for debugging
+                cache_path = Path(hf_constants.HF_HUB_CACHE) / f"models--{repo_id.replace('/', '--')}"
+                print(f"[HF_PATCH] File not cached: {repo_id}/{filename}")
+                print(f"[HF_PATCH] Expected at: {cache_path}")
+            else:
+                print(f"[HF_PATCH] Cache hit: {repo_id}/{filename}")
+            
+            return result
+        
+        # Replace the function
+        import huggingface_hub.file_download as fd
+        fd._try_to_load_from_cache = _patched_try_to_load_from_cache
+        
+        print("[HF_PATCH] huggingface_hub patched for offline mode")
+        
+    except ImportError:
+        print("[HF_PATCH] huggingface_hub not found, skipping patch")
+    except Exception as e:
+        print(f"[HF_PATCH] Error patching huggingface_hub: {e}")
+
+
+def ensure_original_qwen_config_cached():
+    """
+    The MLX community model is based on the original Qwen model.
+    mlx_audio may try to fetch config from the original repo.
+    We need to ensure that config is available in the cache.
+    """
+    from huggingface_hub import constants as hf_constants
+    
+    # Original Qwen model that mlx_audio might reference
+    original_repo = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
+    mlx_repo = "mlx-community/Qwen3-TTS-12Hz-1.7B-Base-bf16"
+    
+    cache_dir = Path(hf_constants.HF_HUB_CACHE)
+    
+    original_path = cache_dir / f"models--{original_repo.replace('/', '--')}"
+    mlx_path = cache_dir / f"models--{mlx_repo.replace('/', '--')}"
+    
+    # If original repo cache doesn't exist but MLX does, create a symlink or copy config
+    if not original_path.exists() and mlx_path.exists():
+        print(f"[HF_PATCH] Original repo not cached, but MLX version is")
+        print(f"[HF_PATCH] Creating symlink from {original_repo} -> {mlx_repo}")
+        
+        try:
+            # Create a symlink so the cache lookup succeeds
+            original_path.parent.mkdir(parents=True, exist_ok=True)
+            original_path.symlink_to(mlx_path, target_is_directory=True)
+            print(f"[HF_PATCH] Symlink created successfully")
+        except Exception as e:
+            print(f"[HF_PATCH] Could not create symlink: {e}")
+
+
+# Auto-apply patch when module is imported
+if os.environ.get("VOICEBOX_OFFLINE_PATCH", "1") != "0":
+    patch_huggingface_hub_offline()
+    ensure_original_qwen_config_cached()

bun.lock

@@ -13,7 +13,7 @@
     },
     "app": {
       "name": "@voicebox/app",
-      "version": "0.1.11",
+      "version": "0.1.13",
       "dependencies": {
         "@dnd-kit/core": "^6.3.1",
         "@dnd-kit/sortable": "^10.0.0",
@@ -68,7 +68,7 @@
     },
     "landing": {
       "name": "@voicebox/landing",
-      "version": "0.1.11",
+      "version": "0.1.13",
       "dependencies": {
         "@radix-ui/react-separator": "^1.1.8",
         "@radix-ui/react-slot": "^1.2.4",
@@ -93,7 +93,7 @@
     },
     "tauri": {
       "name": "@voicebox/tauri",
-      "version": "0.1.11",
+      "version": "0.1.13",
       "dependencies": {
         "@tauri-apps/api": "^2.0.0",
         "@tauri-apps/plugin-dialog": "^2.0.0",
@@ -116,7 +116,7 @@
     },
     "web": {
       "name": "@voicebox/web",
-      "version": "0.1.11",
+      "version": "0.1.13",
       "dependencies": {
         "@tanstack/react-query": "^5.0.0",
         "react": "^18.3.0",
@@ -125,6 +125,7 @@
         "zustand": "^4.5.0",
       },
       "devDependencies": {
+        "@tailwindcss/vite": "^4.0.0",
         "@types/react": "^18.3.0",
         "@types/react-dom": "^18.3.0",
         "@typescript-eslint/eslint-plugin": "^7.0.0",

docker-compose.yml

@@ -0,0 +1,41 @@
+services:
+  voicebox:
+    build: .
+    container_name: voicebox
+    restart: unless-stopped
+
+    ports:
+      # Bind to localhost only for security
+      - "127.0.0.1:17493:17493"
+
+    volumes:
+      # Bind-mount for generated audio (customize the host path as needed)
+      # Host side:      ./output/
+      # Container side: /app/data/generations/
+      - ./output:/app/data/generations
+
+      # Named volume for profiles, DB, cache (persists across container restarts)
+      - voicebox-data:/app/data
+
+      # HuggingFace model cache (so models aren't re-downloaded on rebuild)
+      - huggingface-cache:/home/voicebox/.cache/huggingface
+
+    environment:
+      - LOG_LEVEL=info
+
+    networks:
+      - voicebox-net
+
+    deploy:
+      resources:
+        limits:
+          cpus: '4'
+          memory: 8G
+
+networks:
+  voicebox-net:
+    driver: bridge
+
+volumes:
+  voicebox-data:
+  huggingface-cache:

docs/issue-pain-points.md

@@ -0,0 +1,67 @@
+# Voicebox Issue Pain Points (Snapshot)
+
+## Scope
+
+- Dataset: **128 total issues** (**107 open**, **21 closed**)
+- Source: GitHub issues in `jamiepine/voicebox`
+- Classification: keyword/theme clustering
+- Note: counts below are **non-exclusive** (one issue can belong to multiple pain points)
+
+## Most Common Pain Points (Open Issues)
+
+| Rank | Pain Point | Open Issues | What users are reporting |
+|---|---|---:|---|
+| 1 | Model download & offline reliability | **32** | Downloads failing/stalling, cache/offline behavior inconsistent, wrong model size selected, Errno issues |
+| 2 | GPU/backend compatibility | **22** | GPU not detected, backend fallback surprises, platform-specific runtime failures (Windows/Mac) |
+| 3 | Export/save/file persistence | **15** | Export fails, "failed to fetch/download audio", samples/profiles not saving |
+| 4 | Language/accent quality & coverage | **14** | Missing language support, accent mismatch, robotic outputs |
+| 5 | Update/restart safety + long-op controls | **4** | Auto-restart without warning, update confusion, lack of cancel/pause controls |
+
+## Representative Issues by Pain Point
+
+### 1) Model download & offline reliability (32)
+
+- [#159](https://github.com/jamiepine/voicebox/issues/159) - Qwen download fails with Errno 22
+- [#151](https://github.com/jamiepine/voicebox/issues/151) - Model loading hangs / server crashes
+- [#150](https://github.com/jamiepine/voicebox/issues/150) - Internet required despite downloaded models
+- [#149](https://github.com/jamiepine/voicebox/issues/149) - Cancel/pause controls for large downloads
+- [#96](https://github.com/jamiepine/voicebox/issues/96) - 0.6B selection still uses/downloads 1.7B
+
+### 2) GPU/backend compatibility (22)
+
+- [#164](https://github.com/jamiepine/voicebox/issues/164) - Windows: no GPU usage + multiple breakages
+- [#141](https://github.com/jamiepine/voicebox/issues/141) - Using CPU only, GPU not used
+- [#131](https://github.com/jamiepine/voicebox/issues/131) - Numpy ABI mismatch in bundled app
+- [#130](https://github.com/jamiepine/voicebox/issues/130) - Intel Mac tensor/padding generation error
+- [#127](https://github.com/jamiepine/voicebox/issues/127) - GPU not found
+
+### 3) Export/save/file persistence (15)
+
+- [#148](https://github.com/jamiepine/voicebox/issues/148) - Japanese export fails on 0.1.12
+- [#143](https://github.com/jamiepine/voicebox/issues/143) - Samples not saving
+- [#134](https://github.com/jamiepine/voicebox/issues/134) - Can't save profile
+- [#105](https://github.com/jamiepine/voicebox/issues/105) - Export audio fails (failed to fetch)
+- [#49](https://github.com/jamiepine/voicebox/issues/49) - Export filename/location ignored on Windows
+
+### 4) Language/accent quality & coverage (14)
+
+- [#162](https://github.com/jamiepine/voicebox/issues/162) - Persian audio request/problem
+- [#117](https://github.com/jamiepine/voicebox/issues/117) - Arabic language support
+- [#113](https://github.com/jamiepine/voicebox/issues/113) - Polish language support
+- [#109](https://github.com/jamiepine/voicebox/issues/109) - Ukrainian support
+- [#100](https://github.com/jamiepine/voicebox/issues/100) - Non-US accent quality issues
+
+### 5) Update/restart safety + controls (4)
+
+- [#164](https://github.com/jamiepine/voicebox/issues/164) - Update behavior + usability failures
+- [#136](https://github.com/jamiepine/voicebox/issues/136) - Auto-restart without warning
+- [#86](https://github.com/jamiepine/voicebox/issues/86) - Unexpected restart with no confirmation
+- [#149](https://github.com/jamiepine/voicebox/issues/149) - Need pause/cancel and pre-download confirmation
+
+## Additional Signal
+
+- There is also a large **feature-request/misc** bucket (**36 open**) that is competing with stability triage (audiobook, Linux build, additional ASR/TTS models, integrations).
+
+## Takeaway
+
+Most user pain is concentrated in four stability areas: **download/offline path**, **GPU/backend detection**, **save/export reliability**, and **language/accent correctness**. Addressing those first should reduce the majority of current support friction.

docs/plans/PROJECT_STATUS.md

@@ -1,6 +1,6 @@
 # Voicebox Project Status & Roadmap
 
-> Last updated: 2026-03-12 | Current version: **v0.1.13** | 13.1k stars | 176 open issues | 28 open PRs
+> Last updated: 2026-03-13 | Current version: **v0.1.13** | 13.1k stars | ~176 open issues | 25 open PRs
 
 ---
 
@@ -30,14 +30,18 @@
 │                         │ HTTP :17493                │
 │  ┌──────────────────────▼────────────────────────┐  │
 │  │  FastAPI Backend (backend/)                   │  │
-│  │  ┌─────────────┐  ┌───────────┐  ┌─────────┐ │  │
-│  │  │ TTSBackend  │  │ STTBackend│  │ Profiles│ │  │
-│  │  │ (Protocol)  │  │ (Whisper) │  │ History │ │  │
-│  │  │  ┌────────┐ │  └───────────┘  │ Stories │ │  │
-│  │  │  │PyTorch │ │                  └─────────┘ │  │
-│  │  │  │or MLX  │ │                              │  │
-│  │  │  └────────┘ │                              │  │
-│  │  └─────────────┘                              │  │
+│  │  ┌─────────────────────────────────────────┐  │  │
+│  │  │ TTSBackend Protocol                     │  │  │
+│  │  │  ┌──────────┐ ┌───────┐ ┌───────────┐  │  │  │
+│  │  │  │ Qwen3-TTS│ │LuxTTS │ │Chatterbox │  │  │  │
+│  │  │  │(Py/MLX)  │ │       │ │(MTL+Turbo)│  │  │  │
+│  │  │  └──────────┘ └───────┘ └───────────┘  │  │  │
+│  │  └─────────────────────────────────────────┘  │  │
+│  │  ┌───────────┐  ┌─────────┐                   │  │
+│  │  │ STTBackend│  │ Profiles│                   │  │
+│  │  │ (Whisper) │  │ History │                   │  │
+│  │  └───────────┘  │ Stories │                   │  │
+│  │                  └─────────┘                   │  │
 │  └───────────────────────────────────────────────┘  │
 └─────────────────────────────────────────────────────┘
 ```
@@ -46,131 +50,180 @@
 
 | Layer | File | Purpose |
 |-------|------|---------|
-| Backend entry | `backend/main.py` | FastAPI app, all API routes (~1700 lines) |
+| Backend entry | `backend/main.py` | FastAPI app, all API routes (~2100 lines) |
 | TTS protocol | `backend/backends/__init__.py:14-81` | `TTSBackend` Protocol definition |
-| TTS factory | `backend/backends/__init__.py:118-137` | Singleton backend selection (MLX vs PyTorch) |
+| TTS factory | `backend/backends/__init__.py:138-178` | Thread-safe engine registry (double-checked locking) |
 | PyTorch TTS | `backend/backends/pytorch_backend.py` | Qwen3-TTS via `qwen_tts` package |
 | MLX TTS | `backend/backends/mlx_backend.py` | Qwen3-TTS via `mlx_audio.tts` |
+| LuxTTS | `backend/backends/luxtts_backend.py` | LuxTTS — fast, CPU-friendly |
+| Chatterbox MTL | `backend/backends/chatterbox_backend.py` | Chatterbox Multilingual — 23 languages |
+| Chatterbox Turbo | `backend/backends/chatterbox_turbo_backend.py` | Chatterbox Turbo — English, paralinguistic tags |
 | Platform detect | `backend/platform_detect.py` | Apple Silicon → MLX, else → PyTorch |
 | API types | `backend/models.py` | Pydantic request/response models |
+| HF progress | `backend/utils/hf_progress.py` | HFProgressTracker (tqdm patching for download progress) |
+| Audio utils | `backend/utils/audio.py` | `trim_tts_output()`, normalize, load/save audio |
 | Frontend API | `app/src/lib/api/client.ts` | Hand-written fetch wrapper |
 | Frontend types | `app/src/lib/api/types.ts` | TypeScript API types |
 | Generation form | `app/src/components/Generation/GenerationForm.tsx` | TTS generation UI |
-| Model manager | `app/src/components/ServerSettings/ModelManagement.tsx` | Model download/status UI |
+| Floating gen box | `app/src/components/Generation/FloatingGenerateBox.tsx` | Compact generation UI |
+| Model manager | `app/src/components/ServerSettings/ModelManagement.tsx` | Model download/status/progress UI |
+| GPU acceleration | `app/src/components/ServerSettings/GpuAcceleration.tsx` | CUDA backend swap UI |
 | Gen form hook | `app/src/lib/hooks/useGenerationForm.ts` | Form validation + submission |
+| Language constants | `app/src/lib/constants/languages.ts` | Per-engine language maps |
 
 ### How TTS Generation Works (Current Flow)
 
 ```
 POST /generate
   1. Look up voice profile from DB
-  2. Check model cache → if missing, trigger background download, return HTTP 202
-  3. Load model (lazy): tts_backend.load_model(model_size)
-  4. Create voice prompt: profiles.create_voice_prompt_for_profile()
+  2. Resolve engine from request (qwen | luxtts | chatterbox | chatterbox_turbo)
+  3. Get backend: get_tts_backend_for_engine(engine)  # thread-safe singleton per engine
+  4. Check model cache → if missing, trigger background download, return HTTP 202
+  5. Load model (lazy): tts_backend.load_model(model_size)
+  6. Create voice prompt: profiles.create_voice_prompt_for_profile(engine=engine)
        → tts_backend.create_voice_prompt(audio_path, reference_text)
-  5. Generate: tts_backend.generate(text, voice_prompt, language, seed, instruct)
-  6. Save WAV → data/generations/{id}.wav
-  7. Insert history record in SQLite
-  8. Return GenerationResponse
+  7. Generate: tts_backend.generate(text, voice_prompt, language, seed, instruct)
+  8. Post-process: trim_tts_output() for Chatterbox engines
+  9. Save WAV → data/generations/{id}.wav
+  10. Insert history record in SQLite
+  11. Return GenerationResponse
 ```
 
 ---
 
 ## Current State
 
-### What's Shipped (v0.1.13)
+### What's Shipped (v0.1.13 + recent merges)
 
+**Core TTS:**
 - Qwen3-TTS voice cloning (1.7B and 0.6B models)
 - MLX backend for Apple Silicon, PyTorch for everything else
+- Multi-engine TTS architecture with thread-safe backend registry (PR #254)
+- LuxTTS integration — fast, CPU-friendly English TTS (PR #254)
+- Chatterbox Multilingual TTS — 23 languages including Hebrew (PR #257)
+- Delivery instructions (instruct parameter, Qwen only)
+- Single flat model dropdown (Qwen 1.7B, Qwen 0.6B, LuxTTS, Chatterbox, Chatterbox Turbo)
+
+**Infrastructure:**
+- CUDA backend swap via binary download and restart (PR #252)
+- GPU acceleration settings UI
 - Voice profiles with multi-sample support
 - Stories editor (multi-track DAW timeline)
 - Whisper transcription (base, small, medium, large variants)
-- Model management UI with download progress (SSE)
+- Model management UI with inline download progress bars (HFProgressTracker)
+- Download cancel/clear UI with error panel (PR #238)
 - Generation history with caching
 - Streaming generation endpoint (MLX only)
-- Delivery instructions (instruct parameter)
+- Duplicate profile name validation (PR #175)
+- Linux NVIDIA GBM buffer + WebKitGTK microphone fix (PR #210)
 
-### What's NOT Shipped But Has Code
+### What's In-Flight
 
-| Feature | Branch | Status |
-|---------|--------|--------|
-| External provider binaries (CUDA split) | `external-provider-binaries` | PR #33, significant work done, stale since Feb |
-| Dual server binaries | `feat/dual-server-binaries` | Branch exists, no PR |
-| Multi-sample fix | `fix-multi-sample` | Branch exists, no PR |
-| Model download notification fix | `fix-dl-notification-...` | Branch exists, no PR |
+| Feature | Branch/PR | Status |
+|---------|-----------|--------|
+| Chatterbox Turbo + per-engine language lists | `feat/chatterbox-turbo` / PR #258 | Open, ready for review |
 
-### Hardcoded Qwen3-TTS Assumptions
+### TTS Engine Comparison
 
-These are the specific coupling points that block multi-model support:
+| Engine | Model Name | Languages | Size | Key Features |
+|--------|-----------|-----------|------|-------------|
+| Qwen3-TTS 1.7B | `qwen-tts-1.7B` | 10 (zh, en, ja, ko, de, fr, ru, pt, es, it) | ~3.5 GB | Instruct mode, highest quality |
+| Qwen3-TTS 0.6B | `qwen-tts-0.6B` | 10 | ~1.2 GB | Lighter, faster |
+| LuxTTS | `luxtts` | English | ~300 MB | CPU-friendly, 48 kHz, fast |
+| Chatterbox | `chatterbox-tts` | 23 (incl. Hebrew, Arabic, Hindi, etc.) | ~3.2 GB | Zero-shot cloning, multilingual |
+| Chatterbox Turbo | `chatterbox-turbo` | English | ~1.5 GB | Paralinguistic tags ([laugh], [cough]), 350M params, low latency |
 
-| Location | What's Hardcoded |
-|----------|-----------------|
-| `backend/models.py:58` | `model_size` regex: `^(1\.7B\|0\.6B)$` |
-| `backend/main.py:611` | Default: `model_size or "1.7B"` |
-| `backend/main.py:1322-1365` | Model status list (2 Qwen + 4 Whisper) |
-| `backend/main.py:1523-1548` | Download trigger map |
-| `backend/main.py:1597-1628` | Delete map |
-| `backend/backends/pytorch_backend.py:65-68` | HF repo ID map |
-| `backend/backends/mlx_backend.py:41-44` | MLX repo ID map |
-| `backend/backends/__init__.py:118-137` | Single global TTS backend |
-| `app/src/lib/hooks/useGenerationForm.ts:17` | `modelSize: z.enum(['1.7B', '0.6B'])` |
-| `app/src/lib/hooks/useGenerationForm.ts:70-71` | `modelName = "qwen-tts-${data.modelSize}"` |
-| `app/src/components/Generation/GenerationForm.tsx:140-141` | Hardcoded "Qwen TTS" labels |
-| `app/src/components/ServerSettings/ModelManagement.tsx:166-213` | Filters by `qwen-tts` and `whisper` prefix |
-| `backend/utils/cache.py` | Voice prompt cache uses `torch.save()` |
+### Multi-Engine Architecture (Shipped)
+
+The singleton TTS backend blocker described in the previous version of this doc has been **resolved**. The architecture now supports:
+
+- **Thread-safe backend registry** (`_tts_backends` dict + `_tts_backends_lock`) with double-checked locking
+- **Per-engine backend instances** — each engine gets its own singleton, loaded lazily
+- **Engine field on GenerationRequest** — frontend sends `engine: 'qwen' | 'luxtts' | 'chatterbox' | 'chatterbox_turbo'`
+- **Per-engine language filtering** — `ENGINE_LANGUAGES` map in frontend, backend regex accepts all languages
+- **Per-engine voice prompts** — `create_voice_prompt_for_profile()` dispatches to the correct backend
+- **Trim post-processing** — `trim_tts_output()` for Chatterbox engines (cuts trailing silence/hallucination)
+
+### Known Limitations
+
+- **HF XET progress**: Large files downloaded via `hf-xet` (HuggingFace's new transfer backend) report `n=0` in tqdm updates. Progress bars may appear stuck for large `.safetensors` files even though the download is proceeding. This is a known upstream limitation.
+- **Chatterbox Turbo upstream token bug**: `from_pretrained()` passes `token=os.getenv("HF_TOKEN") or True` which fails without a stored HF token. Our backend works around this by calling `snapshot_download(token=None)` + `from_local()`.
+- **chatterbox-tts must install with `--no-deps`**: It pins `numpy<1.26`, `torch==2.6.0`, `transformers==4.46.3` — all incompatible with our stack (Python 3.12, torch 2.10, transformers 4.57.3). Sub-deps listed explicitly in `requirements.txt`.
+- **Streaming generation** only works for Qwen on MLX. Other engines use the non-streaming `/generate` endpoint.
+- **dicta-onnx** (Hebrew diacritization) not included — upstream Chatterbox bug requires `model_path` arg but calls `Dicta()` with none. Hebrew works fine without it.
 
 ---
 
 ## Open PRs — Triage & Analysis
 
+### Recently Merged (Since Last Update)
+
+| PR | Title | Merged |
+|----|-------|--------|
+| **#257** | feat: Chatterbox TTS engine with multilingual voice cloning | 2026-03-13 |
+| **#254** | feat: LuxTTS integration — multi-engine TTS support | 2026-03-13 |
+| **#252** | feat: CUDA backend swap via binary download and restart | 2026-03-13 |
+| **#238** | Download cancel/clear UI, fixed model downloading | 2026-03-13 |
+| **#250** | docs: align local API port examples | 2026-03-13 |
+| **#210** | fix: Linux NVIDIA GBM buffer crash | 2026-03-13 |
+| **#175** | Fix #134: duplicate profile name validation | 2026-03-13 |
+
+### In-Flight (Our Work)
+
+| PR | Title | Status | Notes |
+|----|-------|--------|-------|
+| **#258** | feat: Chatterbox Turbo engine + per-engine language lists | Open | Ready for review. Adds Turbo engine + dynamic language dropdown. |
+
 ### Merge-Ready / Near-Ready (Bug Fixes & Small Features)
 
 | PR | Title | Risk | Notes |
 |----|-------|------|-------|
-| **#250** | docs: align local API port examples | None | Docs-only |
 | **#230** | docs: fix README grammar | None | Docs-only |
 | **#243** | a11y: screen reader and keyboard improvements | Low | Accessibility, no backend changes |
-| **#175** | Fix #134: duplicate profile name validation | Low | Simple validation |
 | **#178** | Fix #168 #140: generation error handling | Low | Error handling improvements |
 | **#152** | Fix: prevent crashes when HuggingFace unreachable | Medium | Monkey-patches HF hub; solves real offline bug (#150, #151) |
 | **#218** | fix: unify qwen tts cache dir on Windows | Low | Windows-specific path fix |
 | **#214** | fix: panic on launch from tokio::spawn | Low | Rust-side Tauri fix |
-| **#210** | fix: Linux NVIDIA GBM buffer crash | Low | Linux-specific, narrowly scoped |
 | **#88** | security: restrict CORS to known local origins | Low | Security hardening |
+| **#133** | feat: network access toggle | Low | Wires up existing plumbing |
 
 ### Significant Feature PRs
 
-| PR | Title | Complexity | Dependencies | Notes |
-|----|-------|-----------|--------------|-------|
-| **#97** | fix: pass language parameter to TTS models | Medium | None | **Critical bug** — language param was silently dropped. Adds `LANGUAGE_CODE_TO_NAME` mapping to both backends. Should be high priority. |
-| **#133** | feat: network access toggle | Low | None | Wires up existing plumbing (`--host 0.0.0.0`). Clean, small. |
-| **#238** | download cancel/clear UI + error panel | Medium | None | Adds cancel buttons, VS Code-style Problems panel, fixes whisper-large repo. Quality-of-life win. |
-| **#99** | feat: chunked TTS with quality selector | Medium | None | Solves the 500-char/2048-token limit. Sentence-aware splitting, crossfade concat, 44.1kHz upsampling. Addresses #191, #203, #69, #111. |
-| **#154** | feat: Audiobook tab | Medium | Depends on #99 concepts | Full audiobook workflow — chunked gen, preview, auto-save to Stories. New route + tab. |
-| **#91** | fix: CoreAudio device enumeration | Medium | None | macOS audio device handling. |
+| PR | Title | Complexity | Notes |
+|----|-------|-----------|-------|
+| **#253** | Enhance speech tokenizer with 48kHz version | Medium | Qwen tokenizer upgrade |
+| **#97** | fix: pass language parameter to TTS models | Medium | May be partially obsoleted by multi-engine work — needs review |
+| **#99** | feat: chunked TTS with quality selector | Medium | Solves 500-char limit. Addresses #191, #203, #69, #111. |
+| **#154** | feat: Audiobook tab | Medium | Full audiobook workflow. Depends on #99 concepts. |
+| **#91** | fix: CoreAudio device enumeration | Medium | macOS audio device handling |
 
 ### Architectural PRs (Need Careful Review)
 
 | PR | Title | Complexity | Notes |
 |----|-------|-----------|-------|
-| **#33** | CUDA GPU Support — External Provider Binaries | **Very High** | The big one. Splits monolithic backend into main app + downloadable provider executables (PyTorch CPU, CUDA). New provider management system, CI/CD for R2 uploads, provider settings UI. Created Feb 1, significant codebase. **This is the foundation for multi-model support** but is currently Qwen-only. |
-| **#225** | feat: custom HuggingFace model support | High | Adds `custom_models.py`, `custom:<slug>` model IDs, frontend model grouping (Built-in vs Custom). **Takes a different approach than #33** — keeps single backend but allows arbitrary HF repos. These two PRs may conflict architecturally. |
-| **#194** | feat: Hebrew + Chatterbox TTS | High | **First non-Qwen TTS model.** Adds `ChatterboxTTSBackend` alongside existing backends. Routes by language (`he` → Chatterbox, else → Qwen). Adds Hebrew Whisper models. Includes a lot of cleanup. Important precedent for multi-model. |
-| **#195** | feat: per-profile LoRA fine-tuning | **Very High** | Depends on #194. Training pipeline, adapter management, SSE progress, 15 new API endpoints. New DB tables. Forces PyTorch even on MLX systems for adapter inference. |
-| **#161** | feat: Docker + web deployment | High | 3-stage Dockerfile, SPA serving from FastAPI, docker-compose. Implements the Docker deployment plan. |
-| **#124** | Add Dockerfiles + docker-compose + docs | Medium | Earlier, simpler Docker attempt. Overlaps with #161. |
-| **#123** | added docker | Low | Minimal Docker PR. Overlaps with #161 and #124. |
-| **#227** | fix: harden input validation & file safety | Medium | Follow-up to #225. Atomic writes, threading locks, input validation. Good hardening but coupled to the custom models feature. |
+| **#225** | feat: custom HuggingFace model support | High | Arbitrary HF repo loading. May need rework given multi-engine arch is now shipped. |
+| **#194** | feat: Hebrew + Chatterbox TTS | High | **Superseded** by PR #257 which shipped Chatterbox multilingual (23 langs incl. Hebrew). May be closeable. |
+| **#195** | feat: per-profile LoRA fine-tuning | Very High | Training pipeline, adapter management, 15 new endpoints. Depends on #194 (now superseded). |
+| **#161** | feat: Docker + web deployment | High | 3-stage Dockerfile, SPA serving. Independent of TTS engine work. |
+| **#124** / **#123** | Docker (simpler attempts) | Low-Medium | Overlap with #161 |
+| **#227** | fix: harden input validation & file safety | Medium | Coupled to #225 (custom models) |
 
 ### PRs That Need Author Action / Are Stale
 
 | PR | Title | Notes |
 |----|-------|-------|
-| **#237** | fix: bundle qwen_tts source files in PyInstaller | Solves #212 but needs review for build system impact |
+| **#237** | fix: bundle qwen_tts source files in PyInstaller | Build system, needs review |
 | **#215** | Update prerequisites with Tauri deps | Branch is `main` — will have conflicts |
 | **#89** | Linux Support | Branch is `main` — will have conflicts. Broad scope. |
 | **#83** | Update download links for v0.1.12 | Outdated (we're on v0.1.13) |
 
+### PRs Likely Superseded
+
+| PR | Superseded By | Notes |
+|----|--------------|-------|
+| **#194** (Hebrew + Chatterbox) | PR #257 (merged) | #257 ships Chatterbox multilingual with 23 languages including Hebrew. #194 took a different approach (route by language). Can likely be closed. |
+| **#33** (External provider binaries) | PR #252 (merged) | #252 shipped CUDA backend swap. #33's broader provider architecture may still have value but needs reassessment. |
+
 ---
 
 ## Open Issues — Categorized
@@ -186,15 +239,15 @@ The single most reported category. Users on Windows with NVIDIA GPUs frequently
 
 **Key issues:** #239, #222, #220, #217, #208, #198, #192, #167, #164, #141, #130, #127
 
-**Fix path:** PR #33 (external provider binaries) is designed to solve this. Ship a small main app, let users download the CUDA provider separately.
+**Fix path:** PR #252 (CUDA backend swap) is now merged. Users can download the CUDA binary separately from the GPU acceleration settings. Many of these issues may now be resolvable — needs triage to confirm.
 
 ### Model Downloads (20 issues)
 
-Second most reported. Users get stuck downloads, can't resume, no cancel button, no offline fallback.
+Second most reported. Users get stuck downloads, can't resume, no offline fallback.
 
 **Key issues:** #249, #240, #221, #216, #212, #181, #180, #159, #150, #149, #145, #143, #135, #134
 
-**Fix path:** PR #238 (cancel/clear UI), PR #152 (offline crash fix). Resume support not yet addressed.
+**Fix path:** PR #238 (cancel/clear UI) is now merged. PR #152 (offline crash fix) still open. Inline progress bars now show for all engines. Resume support not yet addressed.
 
 ### Language Requests (18 issues)
 
@@ -202,7 +255,7 @@ Strong demand for: Hindi (#245), Indonesian (#247), Dutch (#236), Hebrew (#199),
 
 **Key issues:** #247, #245, #236, #211, #205, #199, #189, #188, #187, #183, #179, #162
 
-**Fix path:** PR #97 (pass language param — currently silently dropped!) is the prerequisite. Qwen3-TTS already supports many languages; the bug is that the language code isn't forwarded. Multi-model (#194 Chatterbox for Hebrew) expands coverage further.
+**Fix path:** Chatterbox Multilingual (merged via #257) now supports 23 languages including many of the requested ones: Arabic, Danish, German, Greek, Finnish, Hebrew, Hindi, Dutch, Norwegian, Polish, Swedish, Swahili, Turkish. Per-engine language filtering (PR #258) ensures the UI shows correct options. Several of these issues may be closeable.
 
 ### New Model Requests (5 explicit issues)
 
@@ -214,7 +267,7 @@ Strong demand for: Hindi (#245), Indonesian (#247), Dutch (#236), Hebrew (#199),
 | #132 | LavaSR (transcription) |
 | #76 | (General model expansion) |
 
-Community is also vocally requesting: LuxTTS, Chatterbox, XTTS-v2, Fish Speech, CosyVoice, Kokoro on social media and in issue comments.
+Community also requests: XTTS-v2, Fish Speech, CosyVoice, Kokoro. The multi-engine architecture is now in place, making new model integration significantly easier.
 
 ### Long-Form / Chunking (5 issues)
 
@@ -255,153 +308,128 @@ Notable requests:
 
 | Document | Target Version | Status | Relevance |
 |----------|---------------|--------|-----------|
-| `TTS_PROVIDER_ARCHITECTURE.md` | v0.1.13 | **Partially implemented** in PR #33 | Core architecture for multi-model + CUDA distribution |
-| `EXTERNAL_PROVIDERS.md` | v0.2.0 | **Not started** | Remote server support. API path inconsistency with provider arch doc (`/v1/` vs `/tts/`) |
-| `MLX_AUDIO.md` | — | **Shipped** (the only one) | MLX backend is live. 0.6B MLX model still missing. |
-| `DOCKER_DEPLOYMENT.md` | v0.2.0 | **PR exists** (#161) | Waiting on review. No official images published. |
-| `OPENAI_SUPPORT.md` | v0.2.0 | **Not started** | OpenAI-compatible API layer. Linked to issue #10. Low complexity. |
-
-### Cross-Document Conflicts
-
-1. **API path inconsistency:** Provider arch uses `/tts/generate`, External providers uses `/v1/generate`, OpenAI compat uses `/v1/audio/speech`. Need to reconcile.
-2. **Docker vs. Provider split:** Docker doc assumes monolithic backend. Provider arch splits into separate binaries. Need to decide: does Docker run the monolith or individual providers?
-3. **Version targeting:** Provider arch targets v0.1.13 (current!) but isn't merged. Everything else targets v0.2.0.
+| `TTS_PROVIDER_ARCHITECTURE.md` | v0.1.13 | **Partially superseded** by multi-engine arch + CUDA swap | Core concepts implemented differently than planned |
+| `CUDA_BACKEND_SWAP.md` | — | **Shipped** (PR #252) | CUDA binary download + backend restart |
+| `CUDA_BACKEND_SWAP_FINAL.md` | — | **Shipped** (PR #252) | Final implementation plan |
+| `EXTERNAL_PROVIDERS.md` | v0.2.0 | **Not started** | Remote server support |
+| `MLX_AUDIO.md` | — | **Shipped** | MLX backend is live |
+| `DOCKER_DEPLOYMENT.md` | v0.2.0 | **PR exists** (#161) | Waiting on review |
+| `OPENAI_SUPPORT.md` | v0.2.0 | **Not started** | OpenAI-compatible API layer |
+| `PR33_CUDA_PROVIDER_REVIEW.md` | — | **Reference** | Analysis of the original provider approach |
 
 ---
 
 ## New Model Integration — Landscape
 
-### Models Worth Supporting (2026 SOTA)
+### Models Worth Supporting (2026 SOTA — updated March 13)
 
-| Model | Cloning | Speed | Sample Rate | Languages | VRAM | Integration Ease | Repo |
-|-------|---------|-------|-------------|-----------|------|-----------------|------|
-| **LuxTTS** | 3s zero-shot | 150x RT, CPU ok | 48 kHz | English-first | <1 GB | Easy | `ysharma3501/LuxTTS` |
-| **Chatterbox** | 5s zero-shot | Sub-200ms streaming | 24-48 kHz | 23+ | Low | Medium | `resemble-ai/chatterbox` |
-| **XTTS-v2** | 6s zero-shot | Fast mid-GPU | 24 kHz | 17+ | Medium | Medium | `coqui/XTTS-v2` |
-| **Fish Speech** | 10-30s few-shot | Real-time | 24-44 kHz | 50+ | Medium | Medium | `fishaudio/fish-speech` |
-| **CosyVoice2-0.5B** | 3-10s zero-shot | Very fast | 24 kHz | Multilingual | Low | Easy | Alibaba HF org |
-| **Kokoro-82M** | 3s instant | CPU realtime | 24 kHz | English | Tiny | Medium | Kokoro repo |
+| Model | Cloning | Speed | Sample Rate | Languages | VRAM | Integration Ease | Status |
+|-------|---------|-------|-------------|-----------|------|-----------------|--------|
+| **Qwen3-TTS** | 10s zero-shot | Medium | 24 kHz | 10 | Medium | **Shipped** | v0.1.13 |
+| **LuxTTS** | 3s zero-shot | 150x RT, CPU ok | 48 kHz | English | <1 GB | **Shipped** | PR #254 |
+| **Chatterbox MTL** | 5s zero-shot | Medium | 24 kHz | 23 | Medium | **Shipped** | PR #257 |
+| **Chatterbox Turbo** | 5s zero-shot | Fast | 24 kHz | English | Low | **PR #258** | In review |
+| **HumeAI TADA 1B/3B** | Zero-shot | 5× faster than LLM-TTS | — | EN (1B), Multilingual (3B) | Medium | Needs vetting | MIT, 700s+ coherent, synced transcript output |
+| **MOSS-TTS Family** | Zero-shot | — | — | Multilingual | Medium | Needs vetting | Apache 2.0, multi-speaker dialogue, text-to-voice design (no ref audio) |
+| **VoxCPM 1.5** | Zero-shot (seconds) | ~0.15 RTF streaming | — | Bilingual (EN/ZH) | Medium | Needs vetting | Apache 2.0, tokenizer-free continuous diffusion, LoRA-friendly |
+| **Pocket TTS** | Zero-shot + streaming | >1× RT on CPU | — | English | ~100M params, CPU-first | Needs vetting | MIT, Kyutai Labs, no GPU required |
+| **Kokoro-82M** | 3s instant | CPU realtime | 24 kHz | English | Tiny (82M) | Ready | Apache 2.0, multi-engine arch in place |
+| **XTTS-v2** | 6s zero-shot | Mid-GPU | 24 kHz | 17+ | Medium | Ready | Multi-engine arch in place |
+| **Fish Speech** | 10-30s few-shot | Real-time | 24-44 kHz | 50+ | Medium | Ready | Multi-engine arch in place |
+| **CosyVoice2-0.5B** | 3-10s zero-shot | Very fast | 24 kHz | Multilingual | Low | Ready | Multi-engine arch in place |
 
-### What's Needed Architecturally for Multi-Model
+#### Notes on New Candidates (March 2026)
 
-The current codebase assumes one TTS model family (Qwen3-TTS). Adding any new model requires:
+- **HumeAI TADA** — Text-Audio Dual Alignment arch. Near-zero hallucinations/drift, free synced transcript. 700+ seconds coherent audio. Best candidate for Stories long-form reliability. [HF: HumeAI/tada-1b](https://huggingface.co/HumeAI/tada-1b) | [GitHub: HumeAI/tada](https://github.com/HumeAI/tada)
+- **MOSS-TTS** — Modular suite: flagship cloning, MOSS-TTSD (multi-speaker dialogue), MOSS-VoiceGenerator (create voices from text descriptions, no ref audio). Unique UX for Stories voice design. [GitHub: OpenMOSS/MOSS-TTS](https://github.com/OpenMOSS/MOSS-TTS)
+- **VoxCPM 1.5** — Tokenizer-free continuous diffusion + autoregressive. No discrete token artifacts. Context-aware prosody/emotion, real-time streaming, LoRA fine-tuning. Trained on 1.8M+ hours. [GitHub: OpenBMB/VoxCPM](https://github.com/OpenBMB/VoxCPM)
+- **Pocket TTS** — 100M param CPU-first model from Kyutai Labs (Moshi team). Runs >1× realtime without GPU. Broadens hardware support significantly. [GitHub: kyutai-labs/pocket-tts](https://github.com/kyutai-labs/pocket-tts)
+- **Watch list:** MioTTS-2.6B (fast LLM-based EN/JP, vLLM compatible), Oolel-Voices (Soynade Research, expressive modular control)
+- **Skipped:** Fish Audio S2 — restrictive research license (commercial use requires approval), despite strong features
 
-1. **Model type concept** — A `model_type` field (e.g. `qwen`, `luxtts`, `chatterbox`) alongside `model_size`. The `GenerationRequest` schema, frontend form, and all model config dicts need updating.
+### Adding a New Engine (Now Straightforward)
 
-2. **Multiple backend instances** — The singleton `get_tts_backend()` needs to become a registry. Different models have different voice prompt formats, different inference APIs, different sample rates.
+With the multi-engine architecture shipped, adding a new TTS engine requires:
 
-3. **Voice prompt format abstraction** — Qwen uses `torch.save()`-serialized tensors. LuxTTS uses `encode_prompt()` returning its own format. Chatterbox uses audio-path-based cloning. The cache system (`backend/utils/cache.py`) needs to handle heterogeneous formats.
+1. **Create `backend/backends/<engine>_backend.py`** — implement `TTSBackend` protocol (~200-300 lines)
+2. **Register in `backend/backends/__init__.py`** — add to `TTS_ENGINES` dict + factory function
+3. **Update `backend/models.py`** — add engine name to regex
+4. **Update `backend/main.py`** — add engine cases in generate, stream, model-status, download, delete (5 dispatch points)
+5. **Update frontend** — add to engine union type, form schema, model dropdown, language map (5-6 files)
 
-4. **Sample rate normalization** — Qwen outputs 24 kHz. LuxTTS outputs 48 kHz. The Stories editor and audio pipeline need to handle mixed rates.
-
-5. **Per-model capabilities** — Not all models support `instruct` (delivery instructions), not all support streaming, not all support the same languages. The UI needs to adapt.
-
-### PR #194 as Precedent
-
-The Hebrew/Chatterbox PR (#194) is the first attempt at multi-model. It takes a pragmatic approach: route by language (`he` → Chatterbox, else → Qwen). This works for one extra model but doesn't scale — what happens when you want Chatterbox for English too?
-
-### PR #225 as Alternative Approach
-
-The custom HuggingFace models PR (#225) takes a different angle: let users register arbitrary HF repos and attempt to load them through the existing Qwen backend. This is flexible but fragile — it assumes all models have the same API as Qwen3-TTS.
-
-### PR #33 as Foundation
-
-The external provider binaries PR (#33) has the most robust architecture for multi-model, since each provider is a separate process with its own dependencies. But it's complex, currently Qwen-only, and has been stale since early February.
+Total effort: **~1 day** for a well-documented model with a PyPI package.
 
 ---
 
 ## Architectural Bottlenecks
 
-### 1. Single Backend Singleton
+### ~~1. Single Backend Singleton~~ — RESOLVED
 
-**File:** `backend/backends/__init__.py:118-137`
+The singleton TTS backend was replaced with a thread-safe per-engine registry in PR #254. Multiple engines can now be loaded simultaneously.
 
-The entire TTS system runs through one global `_tts_backend` instance. You literally cannot have two models loaded. This is the #1 blocker for multi-model support.
+### 2. `main.py` is 2100+ Lines
 
-### 2. `main.py` is 1700+ Lines
+All API routes, all model configs, all business logic in one file. Five separate dispatch points for each engine. Any new engine touches this file in 5 places. A model config registry pattern would reduce duplication.
 
-All API routes, all model configs, all business logic in one file. Three separate hardcoded model config dicts that must stay in sync. Any multi-model change touches this file heavily.
+### 3. Model Config is Scattered (Improved)
 
-### 3. Model Config is Scattered
-
-Model identifiers, HF repo IDs, display names, and download logic are duplicated across:
-- `main.py` (3 separate dicts)
-- `pytorch_backend.py` (HF repo map)
-- `mlx_backend.py` (MLX repo map)
-- `GenerationForm.tsx` (UI labels)
-- `useGenerationForm.ts` (validation schema)
-- `ModelManagement.tsx` (prefix filters)
-
-There is no single source of truth for "what models does Voicebox support."
+Model identifiers are still duplicated across `main.py` (3 dicts), backend files, frontend components, and the languages constant. However, the pattern is now consistent and well-understood. A centralized model registry would help but isn't blocking.
 
 ### 4. Voice Prompt Cache Assumes PyTorch Tensors
 
-`backend/utils/cache.py` uses `torch.save()` / `torch.load()` for caching voice prompts. Models that don't use PyTorch tensors (LuxTTS, MLX-native models) can't use this cache.
+`backend/utils/cache.py` uses `torch.save()` / `torch.load()`. LuxTTS and Chatterbox backends work around this by storing reference audio paths instead of tensors in their voice prompt dicts. Not ideal but functional.
 
-### 5. Frontend Assumes Qwen Model Sizes
+### 5. ~~Frontend Assumes Qwen Model Sizes~~ — RESOLVED
 
-The generation form schema (`useGenerationForm.ts:17`) validates `model_size` as `'1.7B' | '0.6B'`. The model management UI filters by string prefix `qwen-tts`. Adding any model requires touching 3-4 frontend files.
+The generation form now uses a flat model dropdown with engine-based routing. Per-engine language filtering is in place. Model size is only sent for Qwen.
 
 ---
 
 ## Recommended Priorities
 
-### Tier 1 — Ship Now (Bug Fixes & Critical Improvements)
+### Tier 1 — Ship Now (Low Risk)
 
-These PRs fix real user pain with low risk. Can be reviewed and merged quickly.
+| Priority | PR/Item | Impact | Effort |
+|----------|---------|--------|--------|
+| 1 | **#258** — Chatterbox Turbo + per-engine languages | Paralinguistic tags, proper language filtering | Review only |
+| 2 | **#152** — Offline mode crash fix | Fixes #150, #151 | Low |
+| 3 | **#99** — Chunked TTS + quality selector | Removes 500-char limit, addresses 5 issues | Medium |
+| 4 | **#218** — Windows HF cache dir fix | Windows-specific pain | Low |
+| 5 | **#178** — Generation error handling | Error UX | Low |
+| 6 | **#230** — Docs fixes | Zero risk | None |
+| 7 | **#133** — Network access toggle | Wires up existing code | Low |
+| 8 | **#88** — CORS restriction | Security improvement | Low |
+| 9 | **#214** — Tauri window close panic fix | Stability | Low |
+| 10 | Triage GPU issues | Many may be resolved by CUDA swap (#252) | Low |
+| 11 | Close superseded PRs | #194 (superseded by #257), #83 (outdated) | None |
 
-| Priority | PR | Impact | Effort |
-|----------|-----|--------|--------|
-| 1 | **#97** — Pass language param to TTS | Fixes all non-English generation (18 language issues) | Low |
-| 2 | **#238** — Download cancel/clear UI | Addresses 20 download-related issues | Low |
-| 3 | **#152** — Offline mode crash fix | Fixes #150, #151 | Low |
-| 4 | **#99** — Chunked TTS + quality selector | Removes 500-char limit, addresses 5 issues | Medium |
-| 5 | **#218** — Windows HF cache dir fix | Windows-specific pain | Low |
-| 6 | **#175, #178** — Profile validation + error handling | Small fixes | Low |
-| 7 | **#250, #230** — Docs fixes | Zero risk | None |
-| 8 | **#133** — Network access toggle | Wires up existing code | Low |
-| 9 | **#88** — CORS restriction | Security improvement | Low |
-| 10 | **#214** — Tauri window close panic fix | Stability | Low |
+### Tier 2 — Next Release (v0.2.0)
 
-### Tier 2 — Next Release (v0.2.0 Foundations)
-
-These require more review but unlock major capabilities.
-
-| Priority | Item | Impact | Effort | Dependencies |
-|----------|------|--------|--------|-------------|
-| 1 | **PR #33** — External provider binaries | Solves GPU distribution (19 issues), foundation for multi-model | Very High | Needs rebase, thorough review |
-| 2 | **Multi-model abstraction layer** | Required before adding LuxTTS/Chatterbox/etc. | High | Informed by #33, #194, #225 |
-| 3 | **PR #161** — Docker deployment | Server/headless users | Medium | Independent of #33 |
-| 4 | **PR #194** — Hebrew + Chatterbox | First non-Qwen model, language expansion | High | Should align with multi-model abstraction |
-| 5 | **PR #154** — Audiobook tab | Significant feature for long-form users | Medium | Benefits from #99 (chunking) |
+| Priority | Item | Impact | Effort |
+|----------|------|--------|--------|
+| 1 | **#253** — 48kHz speech tokenizer | Quality improvement | Medium |
+| 2 | **#161** — Docker deployment | Server/headless users | Medium |
+| 3 | **#154** — Audiobook tab | Long-form users | Medium |
+| 4 | **Model config registry** | Reduce 5-dispatch-point duplication in main.py | Medium |
+| 5 | **#225** — Custom HuggingFace models | User-supplied models | High (needs rework for multi-engine) |
 
 ### Tier 3 — Future (v0.3.0+)
 
-| Item | Notes |
-|------|-------|
-| LuxTTS integration | 48 kHz, low VRAM, but needs multi-model arch first |
-| XTTS-v2 / Fish Speech | Multilingual powerhouses |
-| OpenAI-compatible API (plan doc exists) | Low effort once API is stable |
-| LoRA fine-tuning (PR #195) | Complex, depends on #194 |
-| External/remote providers (plan doc exists) | Depends on provider architecture |
-| GGUF support (#226) | Depends on model ecosystem maturity |
-| Queue system (#234) | Batch generation |
-| Real-time streaming synthesis | MLX-only currently, needs PyTorch path |
-
-### Decision Point: Multi-Model Architecture
-
-Before adding any new TTS model, a decision is needed on *how*:
-
-**Option A — Provider Binary Split (PR #33 approach)**
-Each model family is a separate executable/process. Most isolated, most flexible, but most complex. Solves the CUDA distribution problem simultaneously.
-
-**Option B — In-Process Model Registry**
-Keep everything in one process but replace the singleton with a registry that can instantiate multiple `TTSBackend` implementations. Simpler, but doesn't solve binary size / CUDA distribution.
-
-**Option C — Hybrid (Recommended)**
-Use Option B for lightweight models (LuxTTS, Kokoro — small, CPU-friendly) that can coexist in-process. Use Option A for heavy models (CUDA Qwen3-TTS, Fish Speech) that need their own process/dependencies. The provider architecture from PR #33 becomes the escape hatch for heavy models, while light models are built-in.
-
-This matches how PR #194 already works (Chatterbox loaded in-process alongside Qwen) while keeping the door open for PR #33's provider split.
+| Priority | Item | Notes |
+|----------|------|-------|
+| 1 | **HumeAI TADA** | Long-form reliability for Stories, synced transcripts. Addresses #234, #203, #191, #111, #69. Needs API vetting. |
+| 2 | **Pocket TTS** (Kyutai) | CPU-first 100M model, broadens hardware support. Kyutai ships clean code. Needs API vetting. |
+| 3 | **MOSS-TTS** | Text-to-voice design (no ref audio) is unique. Multi-speaker dialogue for Stories. Needs thorough API vetting. |
+| 4 | **Kokoro-82M** | 82M params, CPU realtime, Apache 2.0. Easy win. |
+| 5 | **Model config registry refactor** | Reduce 5-dispatch-point duplication in main.py — do before adding 3+ more engines |
+| 6 | XTTS-v2 / Fish Speech / CosyVoice | Multi-engine arch is ready; just needs backend implementation |
+| 7 | **VoxCPM 1.5** | Tokenizer-free streaming, interesting but uncertain integration surface |
+| 8 | OpenAI-compatible API (plan doc exists) | Low effort once API is stable |
+| 9 | LoRA fine-tuning (PR #195) | Complex, needs rework for multi-engine |
+| 10 | External/remote providers | Depends on use case demand |
+| 11 | GGUF support (#226) | Depends on model ecosystem maturity |
+| 12 | Queue system (#234) | Batch generation |
+| 13 | Streaming for non-MLX engines | Currently MLX-only |
 
 ---
 
@@ -409,24 +437,20 @@ This matches how PR #194 already works (Chatterbox loaded in-process alongside Q
 
 | Branch | PR | Status | Notes |
 |--------|-----|--------|-------|
-| `external-provider-binaries` | #33 | Open, stale | Major architecture work |
-| `feat/dual-server-binaries` | — | No PR | Related to provider split? |
+| `feat/chatterbox-turbo` | #258 | Open | Chatterbox Turbo + per-engine languages |
+| `feat/chatterbox` | #257 | **Merged** | Chatterbox Multilingual |
+| `feat/luxtts` | #254 | **Merged** | LuxTTS + multi-engine arch |
+| `external-provider-binaries` | #33 | Superseded by #252 | Original CUDA provider approach |
+| `feat/dual-server-binaries` | — | No PR | Related to provider split |
 | `fix-multi-sample` | — | No PR | Voice profile multi-sample fix |
 | `fix-dl-notification-...` | — | No PR | Model download UX |
-| `improvements` | — | No PR | Unknown scope |
-| `stories` | — | No PR | Stories editor work? |
-| `windows-server-shutdown` | — | No PR | Windows lifecycle |
-| `model-dl-fix` | — | No PR | Model download fix |
-| `channels` | — | No PR | Audio channels |
-| `audio-export-entitlement-fix` | — | No PR | macOS entitlements |
-| `better-docs` | — | No PR | Documentation |
 
 ---
 
 ## Quick Reference: API Endpoints
 
 <details>
-<summary>All current endpoints (v0.1.13)</summary>
+<summary>All current endpoints</summary>
 
 | Endpoint | Method | Purpose |
 |----------|--------|---------|
@@ -437,20 +461,21 @@ This matches how PR #194 already works (Chatterbox loaded in-process alongside Q
 | `/profiles/{id}/avatar` | POST, GET, DELETE | Avatar management |
 | `/profiles/{id}/export` | GET | Export profile as ZIP |
 | `/profiles/import` | POST | Import profile from ZIP |
-| `/generate` | POST | Generate speech |
-| `/generate/stream` | POST | Stream speech (SSE) |
+| `/generate` | POST | Generate speech (engine param selects TTS backend) |
+| `/generate/stream` | POST | Stream speech (MLX only) |
 | `/history` | GET | List generation history |
 | `/history/{id}` | GET, DELETE | Get/delete generation |
 | `/history/{id}/export` | GET | Export generation ZIP |
 | `/history/{id}/export-audio` | GET | Export audio only |
 | `/transcribe` | POST | Transcribe audio (Whisper) |
-| `/models/status` | GET | All model statuses |
+| `/models/status` | GET | All model statuses (Qwen, LuxTTS, Chatterbox, Chatterbox Turbo, Whisper) |
 | `/models/download` | POST | Trigger model download |
+| `/models/download/cancel` | POST | Cancel/dismiss download |
 | `/models/{name}` | DELETE | Delete downloaded model |
 | `/models/load` | POST | Load model into memory |
 | `/models/unload` | POST | Unload model |
 | `/models/progress/{name}` | GET | SSE download progress |
-| `/tasks/active` | GET | Active downloads/generations |
+| `/tasks/active` | GET | Active downloads/generations (with inline progress) |
 | `/stories` | POST, GET | Create/list stories |
 | `/stories/{id}` | GET, PUT, DELETE | Story CRUD |
 | `/stories/{id}/items` | POST, GET | Story items CRUD |
@@ -458,5 +483,8 @@ This matches how PR #194 already works (Chatterbox loaded in-process alongside Q
 | `/channels` | POST, GET | Audio channel CRUD |
 | `/channels/{id}` | PUT, DELETE | Channel update/delete |
 | `/cache/clear` | POST | Clear voice prompt cache |
+| `/server/cuda/status` | GET | CUDA binary availability |
+| `/server/cuda/download` | POST | Download CUDA binary |
+| `/server/cuda/switch` | POST | Switch to CUDA backend |
 
 </details>

tauri/src-tauri/src/audio_capture/linux.rs

@@ -1,16 +1,312 @@
 use crate::audio_capture::AudioCaptureState;
+use base64::{engine::general_purpose, Engine as _};
+use cpal::traits::{DeviceTrait, HostTrait, StreamTrait};
+use cpal::{SampleFormat, StreamConfig};
+use hound::{WavSpec, WavWriter};
+use std::io::Cursor;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::thread;
 
+/// Start capturing system audio on Linux using PulseAudio monitor sources.
+///
+/// PulseAudio exposes "monitor" devices that mirror the output of each sink,
+/// allowing us to capture whatever audio is currently playing on the system.
+/// We use `cpal` with the default host (which will be PulseAudio or PipeWire
+/// on modern Linux) and look for monitor input devices.
 pub async fn start_capture(
     state: &AudioCaptureState,
     max_duration_secs: u32,
 ) -> Result<(), String> {
-    todo!("implement Linux audio capture")
+    // Reset previous samples
+    state.reset();
+
+    let samples = state.samples.clone();
+    let sample_rate_arc = state.sample_rate.clone();
+    let channels_arc = state.channels.clone();
+    let stop_tx = state.stop_tx.clone();
+    let error_arc = state.error.clone();
+
+    // Use AtomicBool for stop signal (works across threads)
+    let stop_flag = Arc::new(AtomicBool::new(false));
+    let stop_flag_clone = stop_flag.clone();
+
+    // Create tokio channel and spawn a task to bridge it to the AtomicBool
+    let (tx, mut rx) = tokio::sync::mpsc::channel::<()>(1);
+    *stop_tx.lock().unwrap() = Some(tx);
+
+    tokio::spawn(async move {
+        rx.recv().await;
+        stop_flag_clone.store(true, Ordering::Relaxed);
+    });
+
+    // Spawn capture on a dedicated thread
+    thread::spawn(move || {
+        let host = cpal::default_host();
+
+        // Try to find a monitor device for system audio capture.
+        // On PulseAudio/PipeWire, monitor sources have "monitor" in their name.
+        let device = {
+            let mut monitor_device = None;
+
+            if let Ok(devices) = host.input_devices() {
+                for d in devices {
+                    if let Ok(name) = d.name() {
+                        let name_lower = name.to_lowercase();
+                        if name_lower.contains("monitor") {
+                            eprintln!("Linux audio capture: Found monitor device: {}", name);
+                            monitor_device = Some(d);
+                            break;
+                        }
+                    }
+                }
+            }
+
+            match monitor_device {
+                Some(d) => d,
+                None => {
+                    // Fallback to default input device (microphone)
+                    eprintln!("Linux audio capture: No monitor device found, falling back to default input");
+                    match host.default_input_device() {
+                        Some(d) => d,
+                        None => {
+                            let error_msg = "No audio input device available".to_string();
+                            eprintln!("{}", error_msg);
+                            *error_arc.lock().unwrap() = Some(error_msg);
+                            return;
+                        }
+                    }
+                }
+            }
+        };
+
+        let device_name = device.name().unwrap_or_else(|_| "unknown".to_string());
+        eprintln!("Linux audio capture: Using device: {}", device_name);
+
+        // Get supported config
+        let config = match device.default_input_config() {
+            Ok(c) => c,
+            Err(e) => {
+                let error_msg = format!("Failed to get default input config: {}", e);
+                eprintln!("{}", error_msg);
+                *error_arc.lock().unwrap() = Some(error_msg);
+                return;
+            }
+        };
+
+        let sample_rate = config.sample_rate().0;
+        let channels = config.channels();
+        let sample_format = config.sample_format();
+
+        eprintln!(
+            "Linux audio capture: Config - {}Hz, {} channels, format: {:?}",
+            sample_rate, channels, sample_format
+        );
+
+        *sample_rate_arc.lock().unwrap() = sample_rate;
+        *channels_arc.lock().unwrap() = channels;
+
+        let stream_config = StreamConfig {
+            channels,
+            sample_rate: cpal::SampleRate(sample_rate),
+            buffer_size: cpal::BufferSize::Default,
+        };
+
+        let samples_clone = samples.clone();
+        let error_arc_clone = error_arc.clone();
+        let stop_flag_for_stream = stop_flag.clone();
+
+        let err_fn = {
+            let error_arc = error_arc.clone();
+            move |err: cpal::StreamError| {
+                let error_msg = format!("Stream error: {}", err);
+                eprintln!("{}", error_msg);
+                *error_arc.lock().unwrap() = Some(error_msg);
+            }
+        };
+
+        let stream = match sample_format {
+            SampleFormat::F32 => {
+                let samples = samples_clone.clone();
+                let stop = stop_flag_for_stream.clone();
+                device.build_input_stream(
+                    &stream_config,
+                    move |data: &[f32], _: &cpal::InputCallbackInfo| {
+                        if stop.load(Ordering::Relaxed) {
+                            return;
+                        }
+                        let mut guard = samples.lock().unwrap();
+                        guard.extend_from_slice(data);
+                    },
+                    err_fn,
+                    None,
+                )
+            }
+            SampleFormat::I16 => {
+                let samples = samples_clone.clone();
+                let stop = stop_flag_for_stream.clone();
+                device.build_input_stream(
+                    &stream_config,
+                    move |data: &[i16], _: &cpal::InputCallbackInfo| {
+                        if stop.load(Ordering::Relaxed) {
+                            return;
+                        }
+                        let mut guard = samples.lock().unwrap();
+                        for &s in data {
+                            guard.push(s as f32 / 32768.0);
+                        }
+                    },
+                    err_fn,
+                    None,
+                )
+            }
+            SampleFormat::U16 => {
+                let samples = samples_clone.clone();
+                let stop = stop_flag_for_stream.clone();
+                device.build_input_stream(
+                    &stream_config,
+                    move |data: &[u16], _: &cpal::InputCallbackInfo| {
+                        if stop.load(Ordering::Relaxed) {
+                            return;
+                        }
+                        let mut guard = samples.lock().unwrap();
+                        for &s in data {
+                            guard.push((s as f32 / 32768.0) - 1.0);
+                        }
+                    },
+                    err_fn,
+                    None,
+                )
+            }
+            _ => {
+                let error_msg = format!("Unsupported sample format: {:?}", sample_format);
+                eprintln!("{}", error_msg);
+                *error_arc_clone.lock().unwrap() = Some(error_msg);
+                return;
+            }
+        };
+
+        let stream = match stream {
+            Ok(s) => s,
+            Err(e) => {
+                let error_msg = format!("Failed to build input stream: {}", e);
+                eprintln!("{}", error_msg);
+                *error_arc_clone.lock().unwrap() = Some(error_msg);
+                return;
+            }
+        };
+
+        if let Err(e) = stream.play() {
+            let error_msg = format!("Failed to start stream: {}", e);
+            eprintln!("{}", error_msg);
+            *error_arc_clone.lock().unwrap() = Some(error_msg);
+            return;
+        }
+
+        eprintln!("Linux audio capture: Stream started successfully");
+
+        // Keep thread alive until stop signal
+        loop {
+            if stop_flag.load(Ordering::Relaxed) {
+                break;
+            }
+            std::thread::sleep(std::time::Duration::from_millis(100));
+        }
+
+        // Stream will be dropped here, stopping capture
+        eprintln!("Linux audio capture: Stream stopped");
+    });
+
+    // Spawn timeout task
+    let stop_tx_clone = state.stop_tx.clone();
+    tokio::spawn(async move {
+        tokio::time::sleep(tokio::time::Duration::from_secs(max_duration_secs as u64)).await;
+        let tx = stop_tx_clone.lock().unwrap().take();
+        if let Some(tx) = tx {
+            let _ = tx.send(()).await;
+        }
+    });
+
+    Ok(())
 }
 
 pub async fn stop_capture(state: &AudioCaptureState) -> Result<String, String> {
-    todo!("implement Linux audio capture stop")
+    // Signal stop
+    if let Some(tx) = state.stop_tx.lock().unwrap().take() {
+        let _ = tx.send(());
+    }
+
+    // Wait a bit for capture to stop
+    tokio::time::sleep(tokio::time::Duration::from_millis(500)).await;
+
+    // Check if there was an error during capture
+    if let Some(error) = state.error.lock().unwrap().as_ref() {
+        return Err(error.clone());
+    }
+
+    // Get samples
+    let samples = state.samples.lock().unwrap().clone();
+    let sample_rate = *state.sample_rate.lock().unwrap();
+    let channels = *state.channels.lock().unwrap();
+
+    if samples.is_empty() {
+        return Err(
+            "No audio samples captured. Make sure audio is playing on your system during recording."
+                .to_string(),
+        );
+    }
+
+    // Convert to WAV
+    let wav_data = samples_to_wav(&samples, sample_rate, channels)?;
+
+    // Encode to base64
+    let base64_data = general_purpose::STANDARD.encode(&wav_data);
+
+    Ok(base64_data)
 }
 
 pub fn is_supported() -> bool {
-    false
+    // Check if we can find a monitor device for system audio capture
+    let host = cpal::default_host();
+    if let Ok(devices) = host.input_devices() {
+        for d in devices {
+            if let Ok(name) = d.name() {
+                if name.to_lowercase().contains("monitor") {
+                    return true;
+                }
+            }
+        }
+    }
+    // Even without a monitor, basic input capture is available
+    host.default_input_device().is_some()
+}
+
+fn samples_to_wav(samples: &[f32], sample_rate: u32, channels: u16) -> Result<Vec<u8>, String> {
+    let mut buffer = Vec::new();
+    let cursor = Cursor::new(&mut buffer);
+
+    let spec = WavSpec {
+        channels,
+        sample_rate,
+        bits_per_sample: 16,
+        sample_format: hound::SampleFormat::Int,
+    };
+
+    let mut writer =
+        WavWriter::new(cursor, spec).map_err(|e| format!("Failed to create WAV writer: {}", e))?;
+
+    // Convert f32 samples to i16
+    for sample in samples {
+        let clamped = sample.clamp(-1.0, 1.0);
+        let i16_sample = (clamped * 32767.0) as i16;
+        writer
+            .write_sample(i16_sample)
+            .map_err(|e| format!("Failed to write sample: {}", e))?;
+    }
+
+    writer
+        .finalize()
+        .map_err(|e| format!("Failed to finalize WAV: {}", e))?;
+
+    Ok(buffer)
 }

tauri/src-tauri/src/main.rs

@@ -16,6 +16,7 @@ struct ServerState {
     child: Mutex<Option<tauri_plugin_shell::process::CommandChild>>,
     server_pid: Mutex<Option<u32>>,
     keep_running_on_close: Mutex<bool>,
+    models_dir: Mutex<Option<String>>,
 }
 
 #[command]
@@ -23,7 +24,16 @@ async fn start_server(
     app: tauri::AppHandle,
     state: State<'_, ServerState>,
     remote: Option<bool>,
+    models_dir: Option<String>,
 ) -> Result<String, String> {
+    // Store models_dir for use on restart (empty string means reset to default)
+    if let Some(ref dir) = models_dir {
+        if dir.is_empty() {
+            *state.models_dir.lock().unwrap() = None;
+        } else {
+            *state.models_dir.lock().unwrap() = Some(dir.clone());
+        }
+    }
     // Check if server is already running (managed by this app instance)
     if state.child.lock().unwrap().is_some() {
         return Ok(format!("http://127.0.0.1:{}", SERVER_PORT));
@@ -274,6 +284,12 @@ async fn start_server(
     let port_str = SERVER_PORT.to_string();
     let is_remote = remote.unwrap_or(false);
 
+    // Resolve the custom models directory from the parameter or stored state
+    let effective_models_dir = models_dir.or_else(|| state.models_dir.lock().unwrap().clone());
+    if let Some(ref dir) = effective_models_dir {
+        println!("Custom models directory: {}", dir);
+    }
+
     // If CUDA binary exists, launch it directly instead of the bundled sidecar
     let spawn_result = if let Some(ref cuda_path) = cuda_binary {
         println!("Launching CUDA backend: {:?}", cuda_path);
@@ -282,6 +298,9 @@ async fn start_server(
         if is_remote {
             cmd = cmd.args(["--host", "0.0.0.0"]);
         }
+        if let Some(ref dir) = effective_models_dir {
+            cmd = cmd.env("VOICEBOX_MODELS_DIR", dir);
+        }
         cmd.spawn()
     } else {
         // Use the bundled CPU sidecar
@@ -289,6 +308,9 @@ async fn start_server(
         if is_remote {
             sidecar = sidecar.args(["--host", "0.0.0.0"]);
         }
+        if let Some(ref dir) = effective_models_dir {
+            sidecar = sidecar.env("VOICEBOX_MODELS_DIR", dir);
+        }
         println!("Spawning server process...");
         sidecar.spawn()
     };
@@ -613,9 +635,19 @@ async fn stop_server(state: State<'_, ServerState>) -> Result<(), String> {
 async fn restart_server(
     app: tauri::AppHandle,
     state: State<'_, ServerState>,
+    models_dir: Option<String>,
 ) -> Result<String, String> {
     println!("restart_server: stopping current server...");
 
+    // Update stored models_dir: empty string means reset to default, non-empty means set
+    if let Some(ref dir) = models_dir {
+        if dir.is_empty() {
+            *state.models_dir.lock().unwrap() = None;
+        } else {
+            *state.models_dir.lock().unwrap() = Some(dir.clone());
+        }
+    }
+
     // Stop the current server
     stop_server(state.clone()).await?;
 
@@ -623,9 +655,9 @@ async fn restart_server(
     println!("restart_server: waiting for port release...");
     tokio::time::sleep(tokio::time::Duration::from_millis(1000)).await;
 
-    // Start server again (will auto-detect CUDA binary)
+    // Start server again (will auto-detect CUDA binary and use stored models_dir)
     println!("restart_server: starting server...");
-    start_server(app, state, None).await
+    start_server(app, state, None, None).await
 }
 
 #[command]
@@ -686,6 +718,7 @@ pub fn run() {
             child: Mutex::new(None),
             server_pid: Mutex::new(None),
             keep_running_on_close: Mutex::new(false),
+            models_dir: Mutex::new(None),
         })
         .manage(audio_capture::AudioCaptureState::new())
         .manage(audio_output::AudioOutputState::new())
@@ -792,7 +825,9 @@ pub fn run() {
                 });
 
                 // Wait for frontend response or timeout
-                tokio::spawn(async move {
+                // Use tauri::async_runtime::spawn instead of tokio::spawn to avoid
+                // panics when the Tokio runtime is being dropped during app shutdown
+                tauri::async_runtime::spawn(async move {
                     tokio::select! {
                         _ = rx.recv() => {
                             // Frontend responded, close window

tauri/src-tauri/tauri.conf.json

@@ -56,7 +56,7 @@
   },
   "plugins": {
     "shell": {
-      "open": true
+      "open": ".*"
     },
     "updater": {
       "pubkey": "dW50cnVzdGVkIGNvbW1lbnQ6IG1pbmlzaWduIHB1YmxpYyBrZXk6IEUxRENBQkRBQjdBNTM1OTIKUldTU05hVzMycXZjNGJGcUxmcVVocll2QjdSaTJNdlFxR2M3VDJsMnVvbDdyZGRPMmRlOW9aWTcK",

tauri/src/platform/lifecycle.ts

@@ -5,9 +5,12 @@ import type { PlatformLifecycle } from '@/platform/types';
 class TauriLifecycle implements PlatformLifecycle {
   onServerReady?: () => void;
 
-  async startServer(remote = false): Promise<string> {
+  async startServer(remote = false, modelsDir?: string | null): Promise<string> {
     try {
-      const result = await invoke<string>('start_server', { remote });
+      const result = await invoke<string>('start_server', {
+        remote,
+        modelsDir: modelsDir ?? undefined,
+      });
       console.log('Server started:', result);
       this.onServerReady?.();
       return result;
@@ -27,9 +30,11 @@ class TauriLifecycle implements PlatformLifecycle {
     }
   }
 
-  async restartServer(): Promise<string> {
+  async restartServer(modelsDir?: string | null): Promise<string> {
     try {
-      const result = await invoke<string>('restart_server');
+      const result = await invoke<string>('restart_server', {
+        modelsDir: modelsDir ?? undefined,
+      });
       console.log('Server restarted:', result);
       this.onServerReady?.();
       return result;

web/src/platform/lifecycle.ts

@@ -3,7 +3,7 @@ import type { PlatformLifecycle } from '@/platform/types';
 class WebLifecycle implements PlatformLifecycle {
   onServerReady?: () => void;
 
-  async startServer(_remote = false): Promise<string> {
+  async startServer(_remote = false, _modelsDir?: string | null): Promise<string> {
     // Web assumes server is running externally
     // Return a default URL - this should be configured via env vars
     const serverUrl = import.meta.env.VITE_SERVER_URL || 'http://localhost:17493';
@@ -15,7 +15,7 @@ class WebLifecycle implements PlatformLifecycle {
     // No-op for web - server is managed externally
   }
 
-  async restartServer(): Promise<string> {
+  async restartServer(_modelsDir?: string | null): Promise<string> {
     // No-op for web - server is managed externally
     return import.meta.env.VITE_SERVER_URL || 'http://localhost:17493';
   }