fix: force offline mode when loading cached models (Qwen TTS & Whisper)

Qwen TTS and Whisper Base make network calls to HuggingFace even when
model weights are fully cached locally, because from_pretrained()
defaults to local_files_only=False. This causes failures for offline
users.

Add a reusable force_offline_if_cached() context manager that sets
HF_HUB_OFFLINE=1 during model loading when is_model_cached() is True.
Applied to all four affected load paths:

- PyTorchTTSBackend (Qwen TTS)
- PyTorchSTTBackend (Whisper)
- MLXTTSBackend (refactored from inline implementation)
- MLXSTTBackend (previously unprotected)

Closes #82

.agents/skills/add-tts-engine/SKILL.md

@@ -0,0 +1,120 @@
+---
+name: add-tts-engine
+description: Use this skill to add a new TTS engine to Voicebox. It walks through dependency research, backend implementation, frontend wiring, PyInstaller bundling, and frozen-build testing. Always start with Phase 0 (dependency audit) before writing any code.
+---
+
+# Add TTS Engine
+
+## Goal
+
+Integrate a new text-to-speech engine into Voicebox end-to-end: dependency research, backend protocol implementation, frontend UI wiring, PyInstaller bundling, and frozen-build verification. The user should only need to test the final build locally.
+
+## Reference Doc
+
+The full phased guide lives at `docs/content/docs/developer/tts-engines.mdx`. **Read this file in its entirety before starting.** It contains:
+
+- Phase 0: Dependency research (mandatory before writing code)
+- Phase 1: Backend implementation (`TTSBackend` protocol)
+- Phase 2: Route and service integration (usually zero changes)
+- Phase 3: Frontend integration (5 files)
+- Phase 4: Dependencies (`requirements.txt`, justfile, CI, Docker)
+- Phase 5: PyInstaller bundling (`build_binary.py` + `server.py`)
+- Phase 6: Common upstream workarounds
+- Implementation checklist (gate between phases)
+
+## Workflow
+
+### 1. Read the guide
+
+```bash
+# Read the full TTS engines doc
+cat docs/content/docs/developer/tts-engines.mdx
+```
+
+Internalize all phases, especially Phase 0 and Phase 5. The v0.2.3 release was three patch releases because Phase 0 was skipped.
+
+### 2. Dependency research (Phase 0)
+
+Clone the model library into a temporary directory and audit it. Do NOT skip this.
+
+```bash
+mkdir /tmp/engine-research && cd /tmp/engine-research
+git clone <model-library-url>
+```
+
+Run the grep searches from Phase 0.2 in the guide against the cloned source and its transitive dependencies. Produce a written dependency audit covering:
+
+1. PyPI vs non-PyPI packages
+2. PyInstaller directives needed (`--collect-all`, `--copy-metadata`, `--hidden-import`)
+3. Runtime data files that must be bundled
+4. Native library paths that need env var overrides in frozen builds
+5. Monkey-patches needed (`torch.load`, float64, MPS, HF token)
+6. Sample rate
+7. Model download method (`from_pretrained` vs `snapshot_download` + `from_local`)
+
+Test model loading and generation on CPU in the throwaway venv before proceeding.
+
+### 3. Implement (Phases 1–4)
+
+Follow the guide's phases in order. Key files to modify:
+
+**Backend (Phase 1):**
+- Create `backend/backends/<engine>_backend.py`
+- Register in `backend/backends/__init__.py` (ModelConfig + TTS_ENGINES + factory)
+- Update regex in `backend/models.py`
+
+**Frontend (Phase 3):**
+- `app/src/lib/api/types.ts` — engine union type
+- `app/src/lib/constants/languages.ts` — ENGINE_LANGUAGES
+- `app/src/components/Generation/EngineModelSelector.tsx` — ENGINE_OPTIONS, ENGINE_DESCRIPTIONS
+- `app/src/lib/hooks/useGenerationForm.ts` — Zod schema, model-name mapping
+- `app/src/components/ServerSettings/ModelManagement.tsx` — MODEL_DESCRIPTIONS
+
+**Dependencies (Phase 4):**
+- `backend/requirements.txt`
+- `justfile` (setup-python, setup-python-release targets)
+- `.github/workflows/release.yml`
+- `Dockerfile` (if applicable)
+
+### 4. PyInstaller bundling (Phase 5)
+
+Register the engine in `backend/build_binary.py`:
+- `--hidden-import` for the backend module and model package
+- `--collect-all` for packages using `inspect.getsource`, shipping data files, or native libraries
+- `--copy-metadata` for packages using `importlib.metadata`
+
+If the engine has native data paths, add `os.environ.setdefault()` in `backend/server.py` inside the `if getattr(sys, 'frozen', False):` block.
+
+### 5. Verify in dev mode
+
+```bash
+just dev
+```
+
+Test the full chain: model download → load → generate → voice cloning.
+
+### 6. Use the checklist
+
+Walk through the Implementation Checklist at the bottom of `tts-engines.mdx`. Every item must be checked before handing the build to the user.
+
+## Key Lessons (from v0.2.3)
+
+These are the most common failure modes. Phase 0 research catches all of them:
+
+| Pattern | Symptom in Frozen Build | Fix |
+|---------|------------------------|-----|
+| `@typechecked` / `inspect.getsource()` | "could not get source code" | `--collect-all <package>` |
+| Package ships pretrained model files | `FileNotFoundError` for `.pth.tar`, `.yaml` | `--collect-all <package>` |
+| C library with hardcoded system paths | `FileNotFoundError` for `/usr/share/...` | `--collect-all` + env var in `server.py` |
+| `importlib.metadata.version()` | "No package metadata found" | `--copy-metadata <package>` |
+| `torch.load` without `map_location` | CUDA device not available on CPU build | Monkey-patch `torch.load` |
+| `torch.from_numpy` on float64 data | dtype mismatch RuntimeError | Cast to `.float()` |
+| `token=True` in HF download calls | Auth failure without stored HF token | Use `snapshot_download(token=None)` + `from_local()` |
+
+## Notes
+
+- The route and service layers have zero per-engine dispatch points. `main.py` requires zero changes.
+- The model config registry in `backends/__init__.py` handles all dispatch automatically.
+- Use `get_torch_device()` and `model_load_progress()` from `backends/base.py` — don't reimplement device detection or progress tracking.
+- Always test with a **clean HuggingFace cache** (no pre-downloaded models from dev).
+- Do NOT push or create a release. Hand the build to the user for local testing.

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.2.3
+current_version = 0.3.1
 commit = True
 tag = True
 tag_name = v{new_version}

.github/workflows/release.yml

@@ -62,6 +62,7 @@ jobs:
           pip install pyinstaller
           pip install -r backend/requirements.txt
           pip install --no-deps chatterbox-tts
+          pip install --no-deps hume-tada
 
       - name: Install MLX dependencies (Apple Silicon only)
         if: matrix.backend == 'mlx'
@@ -188,43 +189,48 @@ jobs:
           pip install pyinstaller
           pip install -r backend/requirements.txt
           pip install --no-deps chatterbox-tts
+          pip install --no-deps hume-tada
 
-      - name: Install PyTorch with CUDA 12.1
+      - name: Install PyTorch with CUDA 12.8
         run: |
-          pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-deps
-          pip install torchaudio --index-url https://download.pytorch.org/whl/cu121
+          pip install torch --index-url https://download.pytorch.org/whl/cu128 --force-reinstall --no-deps
+          pip install torchaudio --index-url https://download.pytorch.org/whl/cu128 --force-reinstall --no-deps
 
       - name: Verify CUDA support in torch
         run: |
           python -c "import torch; print(f'CUDA available in build: {torch.cuda.is_available()}'); print(f'CUDA version: {torch.version.cuda}')"
 
-      - name: Build CUDA server binary
+      - name: Build CUDA server binary (onedir)
         shell: bash
         working-directory: backend
         run: python build_binary.py --cuda
 
-      - name: Split binary for GitHub Releases
+      - name: Package into server core + CUDA libs archives
         shell: bash
         run: |
-          python scripts/split_binary.py \
-            backend/dist/voicebox-server-cuda.exe \
-            --output release-assets/
+          python scripts/package_cuda.py \
+            backend/dist/voicebox-server-cuda/ \
+            --output release-assets/ \
+            --cuda-libs-version cu128-v1 \
+            --torch-compat ">=2.7.0,<2.11.0"
 
-      - name: Upload split parts to GitHub Release
+      - name: Upload archives to GitHub Release
         if: startsWith(github.ref, 'refs/tags/')
-        uses: softprops/action-gh-release@v1
+        uses: softprops/action-gh-release@v2
         with:
           files: |
-            release-assets/voicebox-server-cuda.part*.exe
-            release-assets/voicebox-server-cuda.sha256
-            release-assets/voicebox-server-cuda.manifest
+            release-assets/voicebox-server-cuda.tar.gz
+            release-assets/voicebox-server-cuda.tar.gz.sha256
+            release-assets/cuda-libs-cu128-v1.tar.gz
+            release-assets/cuda-libs-cu128-v1.tar.gz.sha256
+            release-assets/cuda-libs.json
           draft: true
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
-      - name: Upload binary as workflow artifact
+      - name: Upload onedir as workflow artifact
         uses: actions/upload-artifact@v4
         with:
           name: voicebox-server-cuda-windows
-          path: backend/dist/voicebox-server-cuda.exe
+          path: backend/dist/voicebox-server-cuda/
           retention-days: 7

.gitignore

@@ -50,6 +50,14 @@ logs/
 app/openapi.json
 tauri/src-tauri/binaries/*
 tauri/src-tauri/gen/Assets.car
+tauri/src-tauri/gen/voicebox.icns
+tauri/src-tauri/gen/partial.plist
+
+# PyInstaller
+*.spec
+
+# Windows artifacts
+nul
 
 # Temporary
 tmp/

CHANGELOG.md

@@ -7,9 +7,25 @@
 
 ## [Unreleased]
 
-This release rewrites the backend into a modular architecture, migrates the documentation site to Fumadocs, and ships a batch of bug fixes and UI polish across the stack.
+## [0.3.0] - 2026-03-17
 
-The backend's 3,000-line monolith `main.py` has been decomposed into domain routers, a services layer, and a proper database package. A style guide and ruff configuration now enforce consistency. On the frontend, model loading status is now visible in the UI, effects presets get a dropdown, and several race conditions and accessibility gaps are closed.
+This release rewrites the backend into a modular architecture, overhauls the settings UI into routed sub-pages, fixes audio player freezing, migrates documentation to Fumadocs, and ships a batch of bug fixes targeting the most-reported issues from the tracker.
+
+The backend's 3,000-line monolith `main.py` has been decomposed into domain routers, a services layer, and a proper database package. A style guide and ruff configuration now enforce consistency. On the frontend, settings have been split into dedicated routed pages with server logs, a changelog viewer, and an about page. The audio player no longer freezes mid-playback, and model loading status is now visible in the UI. Seven user-reported bugs have been fixed, including server crashes during sample uploads, generation list staleness, cryptic error messages, and CUDA support for RTX 50-series GPUs.
+
+### Settings Overhaul ([#294](https://github.com/jamiepine/voicebox/pull/294))
+- Split settings into routed sub-tabs: General, Generation, GPU, Logs, Changelog, About
+- Added live server log viewer with auto-scroll
+- Added in-app changelog page that parses `CHANGELOG.md` at build time
+- Added About page with version info, license, and generation folder quick-open
+- Extracted reusable `SettingRow` component for consistent setting layouts
+
+### Audio Player Fix ([#293](https://github.com/jamiepine/voicebox/pull/293))
+- Fixed audio player freezing during playback
+- Improved playback UX with better state management and listener cleanup
+- Fixed restart race condition during regeneration
+- Added stable keys for audio element re-rendering
+- Improved accessibility across player controls
 
 ### Backend Refactor ([#285](https://github.com/jamiepine/voicebox/pull/285))
 - Extracted all routes from `main.py` into 13 domain routers under `backend/routes/` — `main.py` dropped from ~3,100 lines to ~10
@@ -40,6 +56,17 @@ The backend's 3,000-line monolith `main.py` has been decomposed into domain rout
 - Softened select focus indicator opacity
 - Addressed 4 critical and 12 major issues from CodeRabbit review
 
+### Bug Fixes ([#295](https://github.com/jamiepine/voicebox/pull/295))
+- Fixed sample uploads crashing the server — audio decoding now runs in a thread pool instead of blocking the async event loop ([#278](https://github.com/jamiepine/voicebox/issues/278))
+- Fixed generation list not updating when a generation completes — switched to `refetchQueries` for reliable cache busting, added SSE error fallback, and page reset on completion ([#231](https://github.com/jamiepine/voicebox/issues/231))
+- Fixed error toasts showing `[object Object]` instead of the actual error message ([#290](https://github.com/jamiepine/voicebox/issues/290))
+- Added Whisper model selection (`base`, `small`, `medium`, `large`, `turbo`) and expanded language support to the `/transcribe` endpoint ([#233](https://github.com/jamiepine/voicebox/issues/233))
+- Upgraded CUDA backend build from cu121 to cu126 for RTX 50-series (Blackwell) GPU support ([#289](https://github.com/jamiepine/voicebox/issues/289))
+- Handled client disconnects in SSE and streaming endpoints to suppress `[Errno 32] Broken Pipe` errors ([#248](https://github.com/jamiepine/voicebox/issues/248))
+- Fixed Docker build failure from pip hash mismatch on Qwen3-TTS dependencies ([#286](https://github.com/jamiepine/voicebox/issues/286))
+- Added 50 MB upload size limit with chunked reads to prevent unbounded memory allocation on sample uploads
+- Eliminated redundant double audio decode in sample processing pipeline
+
 ### Platform Fixes
 - Replaced `netstat` with `TcpStream` + PowerShell for Windows port detection ([#277](https://github.com/jamiepine/voicebox/pull/277))
 - Fixed Docker frontend build and cleaned up Docker docs

Dockerfile

@@ -31,8 +31,12 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential \
     && rm -rf /var/lib/apt/lists/*
 
+RUN pip install --no-cache-dir --upgrade pip
+
 COPY backend/requirements.txt .
 RUN pip install --no-cache-dir --prefix=/install -r requirements.txt
+RUN pip install --no-cache-dir --prefix=/install --no-deps chatterbox-tts
+RUN pip install --no-cache-dir --prefix=/install --no-deps hume-tada
 RUN pip install --no-cache-dir --prefix=/install \
     git+https://github.com/QwenLM/Qwen3-TTS.git
 

README.md

@@ -59,10 +59,10 @@
 
 ## What is Voicebox?
 
-Voicebox is a **local-first voice cloning studio** — a free and open-source alternative to ElevenLabs. Clone voices from a few seconds of audio, generate speech in 23 languages across 4 TTS engines, apply post-processing effects, and compose multi-voice projects with a timeline editor.
+Voicebox is a **local-first voice cloning studio** — a free and open-source alternative to ElevenLabs. Clone voices from a few seconds of audio, generate speech in 23 languages across 5 TTS engines, apply post-processing effects, and compose multi-voice projects with a timeline editor.
 
 - **Complete privacy** — models and voice data stay on your machine
-- **4 TTS engines** — Qwen3-TTS, LuxTTS, Chatterbox Multilingual, and Chatterbox Turbo
+- **5 TTS engines** — Qwen3-TTS, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, and HumeAI TADA
 - **23 languages** — from English to Arabic, Japanese, Hindi, Swahili, and more
 - **Post-processing effects** — pitch shift, reverb, delay, chorus, compression, and filters
 - **Expressive speech** — paralinguistic tags like `[laugh]`, `[sigh]`, `[gasp]` via Chatterbox Turbo
@@ -93,7 +93,7 @@ Voicebox is a **local-first voice cloning studio** — a free and open-source al
 
 ### Multi-Engine Voice Cloning
 
-Four TTS engines with different strengths, switchable per-generation:
+Five TTS engines with different strengths, switchable per-generation:
 
 | Engine                      | Languages | Strengths                                                                                                                                |
 | --------------------------- | --------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
@@ -101,6 +101,7 @@ Four TTS engines with different strengths, switchable per-generation:
 | **LuxTTS**                  | English   | Lightweight (~1GB VRAM), 48kHz output, 150x realtime on CPU                                                                              |
 | **Chatterbox Multilingual** | 23        | Broadest language coverage — Arabic, Danish, Finnish, Greek, Hebrew, Hindi, Malay, Norwegian, Polish, Swahili, Swedish, Turkish and more |
 | **Chatterbox Turbo**        | English   | Fast 350M model with paralinguistic emotion/sound tags                                                                                   |
+| **TADA** (1B / 3B)          | 10        | HumeAI speech-language model — 700s+ coherent audio, text-acoustic dual alignment                                                        |
 
 ### Emotions & Paralinguistic Tags
 
@@ -230,7 +231,7 @@ Full API documentation available at `http://localhost:17493/docs`.
 | Frontend      | React, TypeScript, Tailwind CSS                   |
 | State         | Zustand, React Query                              |
 | Backend       | FastAPI (Python)                                  |
-| TTS Engines   | Qwen3-TTS, LuxTTS, Chatterbox, Chatterbox Turbo   |
+| TTS Engines   | Qwen3-TTS, LuxTTS, Chatterbox, Chatterbox Turbo, TADA |
 | Effects       | Pedalboard (Spotify)                              |
 | Transcription | Whisper / Whisper Turbo (PyTorch or MLX)          |
 | Inference     | MLX (Apple Silicon) / PyTorch (CUDA/ROCm/XPU/CPU) |
@@ -245,7 +246,7 @@ Full API documentation available at `http://localhost:17493/docs`.
 | ----------------------- | ---------------------------------------------- |
 | **Real-time Streaming** | Stream audio as it generates, word by word     |
 | **Voice Design**        | Create new voices from text descriptions       |
-| **More Models**         | XTTS, Bark, and other open-source voice models |
+| **More Models**         | XTTS, Bark, and other open-source voice models  |
 | **Plugin Architecture** | Extend with custom models and effects          |
 | **Mobile Companion**    | Control Voicebox from your phone               |
 
@@ -276,6 +277,12 @@ just build          # Build CPU server binary + Tauri app
 just build-local    # (Windows) Build CPU + CUDA server binaries + Tauri app
 ```
 
+### Adding New Voice Models
+
+The multi-engine architecture makes adding new TTS engines straightforward. A [step-by-step guide](docs/content/docs/developer/tts-engines.mdx) covers the full process: dependency research, backend protocol implementation, frontend wiring, and PyInstaller bundling.
+
+The guide is optimized for AI coding agents. An [agent skill](.agents/skills/add-tts-engine/SKILL.md) can pick up a model name and handle the entire integration autonomously — you just test the build locally.
+
 ### Project Structure
 
 ```

app/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/app",
-  "version": "0.2.3",
+  "version": "0.3.1",
   "private": true,
   "type": "module",
   "scripts": {

app/src/components/AudioPlayer/AudioPlayer.tsx

@@ -17,7 +17,6 @@ export function AudioPlayer() {
     audioUrl,
     audioId,
     profileId,
-    title,
     isPlaying,
     currentTime,
     duration,
@@ -63,7 +62,7 @@ export function AudioPlayer() {
     );
 
     return shouldUseNative;
-  }, [profileChannels, channels, profileId]);
+  }, [profileChannels, channels, platform.metadata.isTauri]);
 
   const waveformRef = useRef<HTMLDivElement>(null);
   const wavesurferRef = useRef<WaveSurfer | null>(null);
@@ -73,31 +72,21 @@ export function AudioPlayer() {
   const isUsingNativePlaybackRef = useRef(false);
   const [isLoading, setIsLoading] = useState(false);
   const [error, setError] = useState<string | null>(null);
+  const [wsReady, setWsReady] = useState(false);
 
-  // Initialize WaveSurfer (only when audioUrl exists and container is ready)
+  // Create WaveSurfer once when the player becomes visible (audioUrl is set).
+  // This instance is reused for all subsequent audio loads - never destroyed until unmount.
   useEffect(() => {
-    // Don't initialize if no audioUrl or already initialized
-    if (!audioUrl) {
-      return;
-    }
+    if (!audioUrl) return;
+    if (wavesurferRef.current) return; // already created
 
-    if (wavesurferRef.current) {
-      debug.log('WaveSurfer already initialized, skipping');
-      return;
-    }
-
-    debug.log('Creating NEW WaveSurfer instance');
-
-    // Wait for container to be properly rendered
     const initWaveSurfer = () => {
       const container = waveformRef.current;
       if (!container) {
-        // Container not ready yet, retry
         setTimeout(initWaveSurfer, 50);
         return;
       }
 
-      // Check if container has dimensions and is visible
       const rect = container.getBoundingClientRect();
       const style = window.getComputedStyle(container);
       const isVisible =
@@ -107,412 +96,221 @@ export function AudioPlayer() {
         style.visibility !== 'hidden';
 
       if (!isVisible) {
-        // Retry after a short delay
         setTimeout(initWaveSurfer, 50);
         return;
       }
 
-      debug.log('Initializing WaveSurfer...', {
-        container,
+      debug.log('Creating WaveSurfer instance', {
         width: rect.width,
         height: rect.height,
       });
 
       try {
-        // Get computed CSS variable values
         const root = document.documentElement;
         const getCSSVar = (varName: string) => {
           const value = getComputedStyle(root).getPropertyValue(varName).trim();
           return value ? `hsl(${value})` : '';
         };
 
-        const waveColor = getCSSVar('--muted');
-        const progressColor = getCSSVar('--accent');
-        const cursorColor = getCSSVar('--accent');
-
         const wavesurfer = WaveSurfer.create({
-          container: container,
-          waveColor: waveColor,
-          progressColor: progressColor,
-          cursorColor: cursorColor,
+          container,
+          waveColor: getCSSVar('--muted'),
+          progressColor: getCSSVar('--accent'),
+          cursorColor: getCSSVar('--accent'),
+          cursorWidth: 3,
           barWidth: 2,
           barRadius: 2,
           height: 80,
           normalize: true,
-          // Use MediaElement backend (default). Unlike the WebAudio backend,
-          // MediaElement uses a standard <audio> element for playback which
-          // benefits from the browser/webview's built-in audio session recovery.
-          // This prevents audio loss when another app steals audio output or
-          // the system audio session is interrupted.
-          interact: true, // Enable interaction (click to seek)
-          mediaControls: false, // Don't show native controls
+          interact: true,
+          dragToSeek: { debounceTime: 0 },
+          mediaControls: false,
+          backend: 'WebAudio',
         });
 
-        wavesurferRef.current = wavesurfer;
-        debug.log('WaveSurfer created successfully');
-      } catch (error) {
-        debug.error('Failed to create WaveSurfer:', error);
-        setError(
-          `Failed to initialize waveform: ${error instanceof Error ? error.message : String(error)}`,
-        );
-        return;
-      }
+        // Wire up event handlers (these persist for the lifetime of the instance)
+        wavesurfer.on('timeupdate', (time) => {
+          const dur = usePlayerStore.getState().duration;
+          if (dur > 0 && time >= dur) {
+            setCurrentTime(dur);
+            const loop = usePlayerStore.getState().isLooping;
+            if (loop) {
+              wavesurfer.seekTo(0);
+              wavesurfer.play().catch((err) => debug.error('Loop play failed:', err));
+            } else {
+              wavesurfer.pause();
+              setIsPlaying(false);
+            }
+            return;
+          }
+          setCurrentTime(time);
+        });
 
-      const wavesurfer = wavesurferRef.current;
-      if (!wavesurfer) return;
+        wavesurfer.on('ready', () => {
+          const dur = wavesurfer.getDuration();
+          setDuration(dur);
+          loadingRef.current = false;
+          setIsLoading(false);
+          setError(null);
+          debug.log('Audio ready, duration:', dur);
 
-      // Update store when time changes, stop if past duration
-      wavesurfer.on('timeupdate', (time) => {
-        const dur = usePlayerStore.getState().duration;
-        if (dur > 0 && time >= dur) {
-          setCurrentTime(dur);
+          wavesurfer.setVolume(usePlayerStore.getState().volume);
+          wavesurfer.setMuted(false);
+
+          // Auto-play if the flag is set (story mode advance or explicit play)
+          const shouldAutoPlayNow = usePlayerStore.getState().shouldAutoPlay;
+          if (shouldAutoPlayNow) {
+            usePlayerStore.getState().clearAutoPlayFlag();
+            wavesurfer.play().catch((err) => {
+              debug.error('Failed to autoplay:', err);
+            });
+          } else {
+            debug.log('Skipping auto-play - shouldAutoPlay is false');
+          }
+        });
+
+        wavesurfer.on('play', () => setIsPlaying(true));
+        wavesurfer.on('pause', () => {
+          setIsPlaying(false);
+          setCurrentTime(wavesurfer.getCurrentTime());
+        });
+
+        wavesurfer.on('seeking', (time) => setCurrentTime(time));
+
+        // Mute audio during drag-to-seek to prevent popping from the WebAudio
+        // backend's hard stop/start cycle on each seek. Unmute with a short
+        // fade-in when the drag ends.
+        const seekMedia = wavesurfer.getMediaElement() as any;
+        const seekGain: GainNode | null = seekMedia?.getGainNode?.() ?? null;
+        if (seekGain) {
+          const ctx = seekGain.context as AudioContext;
+          wavesurfer.on('dragstart', () => {
+            seekGain.gain.cancelScheduledValues(ctx.currentTime);
+            seekGain.gain.setTargetAtTime(0, ctx.currentTime, 0.002);
+          });
+          wavesurfer.on('dragend', () => {
+            seekGain.gain.cancelScheduledValues(ctx.currentTime);
+            seekGain.gain.setTargetAtTime(1, ctx.currentTime, 0.01);
+          });
+        }
+        wavesurfer.on('finish', () => {
           const loop = usePlayerStore.getState().isLooping;
           if (loop) {
             wavesurfer.seekTo(0);
-            wavesurfer.play();
+            wavesurfer.play().catch((err) => debug.error('Loop play failed:', err));
           } else {
-            wavesurfer.pause();
             setIsPlaying(false);
+            const onFinish = usePlayerStore.getState().onFinish;
+            if (onFinish) onFinish();
           }
-          return;
-        }
-        setCurrentTime(time);
-      });
-
-      // Update store when duration is loaded
-      wavesurfer.on('ready', async () => {
-        const dur = wavesurfer.getDuration();
-        setDuration(dur);
-        loadingRef.current = false;
-        setIsLoading(false);
-        setError(null);
-        debug.log('Audio ready, duration:', dur);
-        debug.log('Waveform should be visible now');
-
-        // Ensure volume is set
-        const currentVolume = usePlayerStore.getState().volume;
-        wavesurfer.setVolume(currentVolume);
-
-        // Auto-play when ready - check if we should use native playback
-        // Get current values from the store and queries at runtime (not captured closure values)
-        const currentAudioUrl = usePlayerStore.getState().audioUrl;
-        const currentProfileId = usePlayerStore.getState().profileId;
-
-        debug.log('Auto-play check - capturing runtime values...');
-
-        // Fetch profile channels at runtime (not using captured value)
-        let runtimeProfileChannels = null;
-        let runtimeChannels = null;
-
-        if (platform.metadata.isTauri && currentProfileId) {
-          try {
-            runtimeProfileChannels = await apiClient.getProfileChannels(currentProfileId);
-            debug.log('Runtime profileChannels:', runtimeProfileChannels);
-
-            if (runtimeProfileChannels && runtimeProfileChannels.channel_ids.length > 0) {
-              runtimeChannels = await apiClient.listChannels();
-              debug.log('Runtime channels:', runtimeChannels);
-            }
-          } catch (error) {
-            debug.error('Failed to fetch runtime channel data:', error);
-          }
-        }
-
-        debug.log('Auto-play check:', {
-          isTauri: platform.metadata.isTauri,
-          currentAudioUrl,
-          currentProfileId,
-          hasProfileChannels: !!runtimeProfileChannels,
-          hasChannels: !!runtimeChannels,
         });
 
-        if (
-          platform.metadata.isTauri &&
-          currentAudioUrl &&
-          currentProfileId &&
-          runtimeProfileChannels &&
-          runtimeChannels
-        ) {
-          debug.log('Attempting native audio playback...');
-
-          // Stop any existing native playback first
-          if (isUsingNativePlaybackRef.current) {
-            try {
-              platform.audio.stopPlayback();
-              debug.log('Stopped existing native playback before starting new one');
-            } catch (error) {
-              debug.error('Failed to stop existing playback:', error);
-            }
-          }
-
-          try {
-            // Collect all device IDs from assigned channels
-            const assignedChannels = runtimeChannels.filter((ch: any) =>
-              runtimeProfileChannels.channel_ids.includes(ch.id),
-            );
-            debug.log('Assigned channels for playback:', assignedChannels);
-
-            // Check if any assigned channel has non-default devices
-            const shouldUseNative = assignedChannels.some(
-              (ch: any) => ch.device_ids.length > 0 && !ch.is_default,
-            );
-            debug.log('Should use native playback:', shouldUseNative);
-
-            if (!shouldUseNative) {
-              debug.log('No custom devices assigned, using standard playback');
-              isUsingNativePlaybackRef.current = false;
-            } else {
-              const deviceIds = assignedChannels.flatMap((ch: any) => ch.device_ids);
-              debug.log('Device IDs to play to:', deviceIds);
-
-              if (deviceIds.length > 0) {
-                debug.log('Fetching audio data from:', currentAudioUrl);
-                // Fetch audio data
-                const response = await fetch(currentAudioUrl);
-                const audioData = new Uint8Array(await response.arrayBuffer());
-                debug.log('Audio data size:', audioData.length);
-
-                // Play via native audio
-                debug.log('Invoking play_audio_to_devices...');
-                try {
-                  await platform.audio.playToDevices(audioData, deviceIds);
-                  debug.log('play_audio_to_devices completed successfully');
-
-                  // Mark that we're using native playback
-                  isUsingNativePlaybackRef.current = true;
-
-                  // Mute WaveSurfer's audio output — native handles the actual sound
-                  // Keep WaveSurfer running for waveform visualization
-                  wavesurfer.setVolume(0);
-                  wavesurfer.setMuted(true);
-
-                  // Start WaveSurfer playback for visualization (muted)
-                  wavesurfer.play().catch((error) => {
-                    debug.error('Failed to start WaveSurfer visualization:', error);
-                  });
-
-                  setIsPlaying(true);
-                  debug.log('Auto-playing via native audio routing - SUCCESS');
-                  return;
-                } catch (invokeError) {
-                  debug.error('play_audio_to_devices invoke failed:', invokeError);
-                  throw invokeError;
-                }
-              } else {
-                debug.log('No device IDs found, falling back to WaveSurfer');
-              }
-            }
-          } catch (error) {
-            debug.error(
-              'Native playback failed during auto-play, falling back to WaveSurfer:',
-              error,
-            );
-            isUsingNativePlaybackRef.current = false;
-            // Fall through to WaveSurfer playback
-          }
-        }
-
-        // Standard playback path — ensure WaveSurfer is unmuted
-        if (!isUsingNativePlaybackRef.current) {
-          wavesurfer.setMuted(false);
-          wavesurfer.setVolume(usePlayerStore.getState().volume);
-        }
-
-        // Only auto-play if shouldAutoPlay flag is set (user explicitly clicked to play)
-        const shouldAutoPlayNow = usePlayerStore.getState().shouldAutoPlay;
-        if (shouldAutoPlayNow) {
-          // Clear the flag first
-          usePlayerStore.getState().clearAutoPlayFlag();
-
-          // Use a small delay to ensure audio element is fully ready
-          setTimeout(() => {
-            wavesurfer.play().catch((error) => {
-              debug.error('Failed to autoplay:', error);
-              // Don't show error for autoplay failures (browser restrictions)
-            });
-          }, 100);
-        } else {
-          debug.log('Skipping auto-play - shouldAutoPlay is false');
-        }
-      });
-
-      // Handle play/pause
-      wavesurfer.on('play', () => {
-        setIsPlaying(true);
-      });
-      wavesurfer.on('pause', () => setIsPlaying(false));
-      wavesurfer.on('finish', () => {
-        // Check loop state from store
-        const loop = usePlayerStore.getState().isLooping;
-        if (loop) {
-          wavesurfer.seekTo(0);
-          wavesurfer.play();
-        } else {
-          setIsPlaying(false);
-          // Trigger finish callback if set
-          const onFinish = usePlayerStore.getState().onFinish;
-          if (onFinish) {
-            onFinish();
-          }
-        }
-      });
-
-      // Handle errors
-      wavesurfer.on('error', (error) => {
-        debug.error('WaveSurfer error:', error);
-        setIsLoading(false);
-        setError(`Audio error: ${error instanceof Error ? error.message : String(error)}`);
-      });
-
-      // Handle loading
-      wavesurfer.on('loading', (percent) => {
-        setIsLoading(true);
-        if (percent === 100) {
+        wavesurfer.on('error', (err) => {
+          debug.error('WaveSurfer error:', err);
           setIsLoading(false);
-        }
-      });
+          setError(`Audio error: ${err instanceof Error ? err.message : String(err)}`);
+        });
 
-      // Load audio immediately if audioUrl is already set
-      if (audioUrl) {
-        debug.log('WaveSurfer ready, loading audio:', audioUrl);
-        loadingRef.current = true;
-        setIsLoading(true);
-        // Stop any current playback before loading new audio
-        if (wavesurfer.isPlaying()) {
-          wavesurfer.pause();
-        }
-        wavesurfer
-          .load(audioUrl)
-          .then(() => {
-            debug.log('Audio loaded into WaveSurfer');
-            loadingRef.current = false;
-          })
-          .catch((error) => {
-            debug.error('Failed to load audio into WaveSurfer:', error);
-            loadingRef.current = false;
-            setIsLoading(false);
-            setError(
-              `Failed to load audio: ${error instanceof Error ? error.message : String(error)}`,
-            );
-          });
+        wavesurfer.on('loading', (percent) => {
+          setIsLoading(true);
+          if (percent === 100) setIsLoading(false);
+        });
+
+        wavesurferRef.current = wavesurfer;
+        setWsReady(true);
+        debug.log('WaveSurfer created successfully');
+      } catch (err) {
+        debug.error('Failed to create WaveSurfer:', err);
+        setError(
+          `Failed to initialize waveform: ${err instanceof Error ? err.message : String(err)}`,
+        );
       }
     };
 
-    // Use double requestAnimationFrame to ensure DOM is fully rendered
-    let rafId1: number;
-    let rafId2: number;
-    let timeoutId: number | null = null;
-
-    rafId1 = requestAnimationFrame(() => {
-      rafId2 = requestAnimationFrame(() => {
-        // Add a small delay to ensure container is fully laid out
-        timeoutId = setTimeout(() => {
-          initWaveSurfer();
-        }, 10);
-      });
+    let rafId: number;
+    rafId = requestAnimationFrame(() => {
+      initWaveSurfer();
     });
 
     return () => {
-      debug.log('Cleaning up WaveSurfer initialization effect');
-      if (rafId1) cancelAnimationFrame(rafId1);
-      if (rafId2) cancelAnimationFrame(rafId2);
-      if (timeoutId) clearTimeout(timeoutId);
+      cancelAnimationFrame(rafId);
+    };
+    // Only run on mount-like conditions. audioUrl is here so we create the instance
+    // when the player first appears, but we guard against re-creation above.
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [audioUrl, setIsPlaying, setDuration, setCurrentTime]);
+
+  // Destroy WaveSurfer only on unmount
+  useEffect(() => {
+    return () => {
       if (wavesurferRef.current) {
-        debug.log('Destroying WaveSurfer instance');
+        debug.log('Destroying WaveSurfer instance (unmount)');
         try {
           wavesurferRef.current.destroy();
-        } catch (error) {
-          debug.error('Error destroying WaveSurfer:', error);
+        } catch (err) {
+          debug.error('Error destroying WaveSurfer:', err);
         }
         wavesurferRef.current = null;
+        setWsReady(false);
       }
     };
-  }, [audioUrl, setIsPlaying, setCurrentTime, setDuration]);
+  }, []);
 
-  // Load audio when URL changes (only if WaveSurfer is already initialized)
+  // Load audio when URL changes (reuses the existing WaveSurfer instance)
   useEffect(() => {
     const wavesurfer = wavesurferRef.current;
+    if (!wavesurfer || !wsReady) return;
 
-    if (!audioUrl || !wavesurfer) {
-      // Reset state when no audio or WaveSurfer not ready
-      if (!audioUrl && wavesurfer) {
-        wavesurfer.pause();
-        wavesurfer.seekTo(0);
-        loadingRef.current = false;
-        setIsLoading(false);
-        setDuration(0);
-        setCurrentTime(0);
-        setError(null);
-        // Reset native playback flag
-        isUsingNativePlaybackRef.current = false;
-      }
+    if (!audioUrl) {
+      // No audio - pause and reset
+      wavesurfer.pause();
+      wavesurfer.seekTo(0);
+      loadingRef.current = false;
+      setIsLoading(false);
+      setDuration(0);
+      setCurrentTime(0);
+      setError(null);
+      isUsingNativePlaybackRef.current = false;
       return;
     }
 
-    // Stop native playback if it was active
-    if (isUsingNativePlaybackRef.current && platform.metadata.isTauri) {
-      try {
-        platform.audio.stopPlayback();
-        debug.log('Stopped native audio playback');
-      } catch (error) {
-        debug.error('Failed to stop native playback:', error);
-      }
-    }
-
-    // Reset native playback flag when loading new audio
-    // Unmute WaveSurfer if it was muted for native playback
-    if (isUsingNativePlaybackRef.current) {
-      wavesurfer.setMuted(false);
-      wavesurfer.setVolume(usePlayerStore.getState().volume);
-    }
+    // Reset native playback state
     isUsingNativePlaybackRef.current = false;
+    wavesurfer.setMuted(false);
+    wavesurfer.setVolume(usePlayerStore.getState().volume);
 
-    // CRITICAL: Force stop any current playback and cancel any pending loads
-    // This must happen BEFORE any early returns
-    debug.log('Audio URL changed to:', audioUrl);
-
-    // COMPLETELY stop and destroy the current audio
+    // Stop current playback and reset position before loading new audio.
+    // With the WebAudio backend, pause() accumulates playedDuration internally.
+    // seekTo(0) resets it so the new track starts from the beginning.
+    debug.log('Loading new audio URL:', audioUrl);
     try {
-      // First pause if playing
       if (wavesurfer.isPlaying()) {
-        debug.log('Pausing current playback');
         wavesurfer.pause();
       }
-
-      // Use empty() to completely destroy the waveform and reset media
-      debug.log('Calling wavesurfer.empty() to destroy audio');
-      wavesurfer.empty();
-    } catch (error) {
-      debug.error('Error stopping previous audio:', error);
-      // Continue anyway to load new audio
+      wavesurfer.seekTo(0);
+    } catch (err) {
+      debug.error('Error resetting before load:', err);
     }
 
-    // Reset loading state to allow new load (cancel any pending loads)
-    loadingRef.current = false;
-
-    // Now start the new load
     loadingRef.current = true;
     setIsLoading(true);
     setError(null);
     setCurrentTime(0);
     setDuration(0);
 
-    // Load new audio
-    debug.log('Starting new audio load for:', audioUrl);
     wavesurfer
       .load(audioUrl)
       .then(() => {
-        debug.log('Audio load promise resolved');
-        // Don't set loading to false here - wait for 'ready' event
+        debug.log('Audio loaded into WaveSurfer');
+        loadingRef.current = false;
       })
-      .catch((error) => {
-        debug.error('Failed to load audio:', error);
-        debug.error('Audio URL:', audioUrl);
+      .catch((err) => {
+        debug.error('Failed to load audio:', err);
         loadingRef.current = false;
         setIsLoading(false);
-        setError(`Failed to load audio: ${error instanceof Error ? error.message : String(error)}`);
+        setError(`Failed to load audio: ${err instanceof Error ? err.message : String(err)}`);
       });
-  }, [audioUrl, setCurrentTime, setDuration]);
+  }, [audioUrl, wsReady, setCurrentTime, setDuration]);
 
   // Sync play/pause state (only when user clicks play/pause button, not auto-sync)
   // This effect is kept for external state changes but should be minimal
@@ -520,7 +318,6 @@ export function AudioPlayer() {
     if (!wavesurferRef.current || duration === 0) return;
 
     if (isPlaying && wavesurferRef.current.isPlaying() === false) {
-      // Only auto-play if audio is ready
       wavesurferRef.current.play().catch((error) => {
         debug.error('Failed to play:', error);
         setIsPlaying(false);
@@ -534,14 +331,7 @@ export function AudioPlayer() {
   // Sync volume
   useEffect(() => {
     if (wavesurferRef.current) {
-      // If using native playback, keep WaveSurfer muted regardless of volume setting
-      if (isUsingNativePlaybackRef.current) {
-        wavesurferRef.current.setVolume(0);
-        debug.log('Volume sync: Using native playback, keeping WaveSurfer muted');
-      } else {
-        wavesurferRef.current.setVolume(volume);
-        debug.log('Volume synced:', volume);
-      }
+      wavesurferRef.current.setVolume(volume);
     }
   }, [volume]);
 
@@ -566,7 +356,6 @@ export function AudioPlayer() {
       return;
     }
 
-    // Reset to beginning and play
     debug.log('Restarting current audio from beginning');
     wavesurfer.seekTo(0);
     wavesurfer.play().catch((error) => {
@@ -575,34 +364,35 @@ export function AudioPlayer() {
       setError(`Playback error: ${error instanceof Error ? error.message : String(error)}`);
     });
 
-    // Clear the restart flag
     clearRestartFlag();
   }, [shouldRestart, duration, setIsPlaying, clearRestartFlag]);
 
-  // Handle shouldAutoPlay flag - for story mode auto-advance
-  const shouldAutoPlay = usePlayerStore((state) => state.shouldAutoPlay);
-  const clearAutoPlayFlag = usePlayerStore((state) => state.clearAutoPlayFlag);
+  // Auto-play is handled exclusively in the WaveSurfer 'ready' event handler.
+  // A separate effect here would race with the ready event since the WebAudio
+  // backend needs to fully decode the audio before play() works correctly.
 
+  // Spacebar to play/pause (capture phase so it fires before focused elements)
   useEffect(() => {
-    const wavesurfer = wavesurferRef.current;
-    if (!wavesurfer || !shouldAutoPlay || duration === 0) {
-      return;
-    }
-
-    // Auto-play the newly loaded audio
-    debug.log('Auto-playing next track in story mode');
-    wavesurfer.seekTo(0);
-    wavesurfer.play().catch((error) => {
-      debug.error('Failed to auto-play:', error);
-      setIsPlaying(false);
-      setError(`Playback error: ${error instanceof Error ? error.message : String(error)}`);
-    });
-
-    // Clear the auto-play flag
-    clearAutoPlayFlag();
-  }, [shouldAutoPlay, duration, setIsPlaying, clearAutoPlayFlag]);
-
-  // Handle loop - WaveSurfer handles this via the 'finish' event
+    const onKeyDown = (e: KeyboardEvent) => {
+      if (e.code !== 'Space') return;
+      // Ignore if user is typing in an input/textarea
+      const tag = (e.target as HTMLElement)?.tagName;
+      if (tag === 'INPUT' || tag === 'TEXTAREA' || (e.target as HTMLElement)?.isContentEditable) {
+        return;
+      }
+      if (audioUrl && duration > 0 && wavesurferRef.current) {
+        e.preventDefault();
+        e.stopPropagation();
+        if (wavesurferRef.current.isPlaying()) {
+          wavesurferRef.current.pause();
+        } else {
+          wavesurferRef.current.play().catch((err) => debug.error('Spacebar play failed:', err));
+        }
+      }
+    };
+    document.addEventListener('keydown', onKeyDown, true);
+    return () => document.removeEventListener('keydown', onKeyDown, true);
+  }, [audioUrl, duration]);
 
   const handlePlayPause = async () => {
     // Standard WaveSurfer playback (works for both normal and native playback modes)
@@ -741,32 +531,32 @@ export function AudioPlayer() {
             size="icon"
             onClick={handlePlayPause}
             disabled={isLoading || duration === 0}
-            className="shrink-0"
+            className={`shrink-0 -mt-2 ${isPlaying ? 'bg-accent text-accent-foreground' : ''}`}
             title={duration === 0 && !isLoading ? 'Audio not loaded' : ''}
             aria-label={
               duration === 0 && !isLoading ? 'Audio not loaded' : isPlaying ? 'Pause' : 'Play'
             }
           >
-            {isPlaying ? <Pause className="h-5 w-5" /> : <Play className="h-5 w-5" />}
+            {isPlaying ? (
+              <Pause className="h-5 w-5 fill-current" />
+            ) : (
+              <Play className="h-5 w-5 fill-current" />
+            )}
           </Button>
 
           {/* Waveform */}
           <div className="flex-1 min-w-0 flex flex-col gap-1">
-            <div ref={waveformRef} className="w-full min-h-[80px]" />
-            {duration > 0 && (
-              <Slider
-                value={duration > 0 ? [(currentTime / duration) * 100] : [0]}
-                onValueChange={handleSeek}
-                max={100}
-                step={0.1}
-                className="w-full"
-                aria-label="Playback position"
-                aria-valuetext={`${formatAudioDuration(currentTime)} of ${formatAudioDuration(duration)}`}
-              />
-            )}
-            {isLoading && (
-              <div className="text-xs text-muted-foreground text-center py-2">Loading audio...</div>
-            )}
+            <div ref={waveformRef} className="w-full min-h-[80px] select-none" />
+            <Slider
+              value={duration > 0 ? [(currentTime / duration) * 100] : [0]}
+              onValueChange={handleSeek}
+              max={100}
+              step={0.1}
+              className="w-full"
+              aria-label="Playback position"
+              aria-valuetext={`${formatAudioDuration(currentTime)} of ${formatAudioDuration(duration)}`}
+            />
+
             {error && <div className="text-xs text-destructive text-center py-2">{error}</div>}
           </div>
 
@@ -777,19 +567,12 @@ export function AudioPlayer() {
             <span className="font-mono">{formatAudioDuration(duration)}</span>
           </div>
 
-          {/* Title */}
-          {title && (
-            <div className="text-sm font-medium truncate max-w-[200px] shrink-0 hidden lg:block">
-              {title}
-            </div>
-          )}
-
           {/* Loop Button */}
           <Button
             variant="ghost"
             size="icon"
             onClick={toggleLoop}
-            className={isLooping ? 'text-primary' : ''}
+            className={isLooping ? 'bg-accent text-accent-foreground' : ''}
             title="Toggle loop"
             aria-label={isLooping ? 'Stop looping' : 'Loop'}
           >

app/src/components/Generation/EngineModelSelector.tsx

@@ -20,6 +20,8 @@ const ENGINE_OPTIONS = [
   { value: 'luxtts', label: 'LuxTTS' },
   { value: 'chatterbox', label: 'Chatterbox' },
   { value: 'chatterbox_turbo', label: 'Chatterbox Turbo' },
+  { value: 'tada:1B', label: 'TADA 1B' },
+  { value: 'tada:3B', label: 'TADA 3B Multilingual' },
 ] as const;
 
 const ENGINE_DESCRIPTIONS: Record<string, string> = {
@@ -27,6 +29,7 @@ const ENGINE_DESCRIPTIONS: Record<string, string> = {
   luxtts: 'Fast, English-focused',
   chatterbox: '23 languages, incl. Hebrew',
   chatterbox_turbo: 'English, [laugh] [cough] tags',
+  tada: 'HumeAI, 700s+ coherent audio',
 };
 
 /** Engines that only support English and should force language to 'en' on select. */
@@ -34,6 +37,7 @@ const ENGLISH_ONLY_ENGINES = new Set(['luxtts', 'chatterbox_turbo']);
 
 function getSelectValue(engine: string, modelSize?: string): string {
   if (engine === 'qwen') return `qwen:${modelSize || '1.7B'}`;
+  if (engine === 'tada') return `tada:${modelSize || '1B'}`;
   return engine;
 }
 
@@ -48,6 +52,20 @@ function handleEngineChange(form: UseFormReturn<GenerationFormValues>, value: st
     if (!available.some((l) => l.value === currentLang)) {
       form.setValue('language', available[0]?.value ?? 'en');
     }
+  } else if (value.startsWith('tada:')) {
+    const [, modelSize] = value.split(':');
+    form.setValue('engine', 'tada');
+    form.setValue('modelSize', modelSize as '1B' | '3B');
+    // TADA 1B is English-only; 3B is multilingual
+    if (modelSize === '1B') {
+      form.setValue('language', 'en');
+    } else {
+      const currentLang = form.getValues('language');
+      const available = getLanguageOptionsForEngine('tada');
+      if (!available.some((l) => l.value === currentLang)) {
+        form.setValue('language', available[0]?.value ?? 'en');
+      }
+    }
   } else {
     form.setValue('engine', value as GenerationFormValues['engine']);
     form.setValue('modelSize', undefined as unknown as '1.7B' | '0.6B');

app/src/components/History/HistoryTable.tsx

@@ -64,7 +64,7 @@ function AudioBars({ mode }: { mode: 'idle' | 'generating' | 'playing' }) {
     <div className="flex items-center gap-[2px] h-5">
       {[0, 1, 2, 3, 4].map((i) => (
         <motion.div
-          key={i}
+          key={`${mode}-${i}`}
           className={`w-[3px] rounded-full ${barColor}`}
           animate={
             mode === 'generating'
@@ -153,7 +153,9 @@ export function HistoryTable() {
     }
   }, [historyData, page]);
 
-  // Reset to page 0 when deletions or imports occur
+  // Reset to page 0 when deletions, imports, or generation completions occur
+  const pendingCount = useGenerationStore((state) => state.pendingGenerationIds.size);
+  const prevPendingCountRef = useRef(pendingCount);
   useEffect(() => {
     if (deleteGeneration.isSuccess || importGeneration.isSuccess) {
       setPage(0);
@@ -161,6 +163,19 @@ export function HistoryTable() {
     }
   }, [deleteGeneration.isSuccess, importGeneration.isSuccess]);
 
+  useEffect(() => {
+    // A generation finished (pending count decreased) — scroll back to show it
+    if (
+      prevPendingCountRef.current > 0 &&
+      pendingCount < prevPendingCountRef.current &&
+      page !== 0
+    ) {
+      setPage(0);
+      setAllHistory([]);
+    }
+    prevPendingCountRef.current = pendingCount;
+  }, [pendingCount, page]);
+
   // Intersection Observer for infinite scroll
   useEffect(() => {
     const loadMoreEl = loadMoreRef.current;
@@ -440,7 +455,7 @@ export function HistoryTable() {
                     role={isPlayable ? 'button' : undefined}
                     tabIndex={isPlayable ? 0 : undefined}
                     className={cn(
-                      'flex items-stretch gap-4 h-26 p-3',
+                      'flex items-stretch gap-4 h-26 p-3 outline-none',
                       isPlayable && 'hover:bg-muted/70 cursor-pointer rounded-md',
                       isVersionsExpanded && 'rounded-b-none',
                     )}

app/src/components/ServerSettings/GpuAcceleration.tsx

@@ -243,7 +243,40 @@ export function GpuAcceleration() {
 
         {/* Native GPU detected - no CUDA download needed */}
 
-        {/* CUDA download section - only show when no GPU is active (native or CUDA) */}
+        {/* Currently running CUDA - show switch back to CPU */}
+        {isCurrentlyCuda && platform.metadata.isTauri && (
+          <>
+            {restartPhase !== 'idle' ? (
+              <div className="flex items-center gap-2 p-3 rounded-lg bg-primary/5 border">
+                <Loader2 className="h-4 w-4 animate-spin" />
+                <span className="text-sm">
+                  {restartPhase === 'stopping' && 'Stopping server...'}
+                  {restartPhase === 'waiting' && 'Restarting server...'}
+                  {restartPhase === 'ready' && 'Server restarted successfully!'}
+                </span>
+              </div>
+            ) : (
+              <div className="space-y-3">
+                <p className="text-sm text-muted-foreground">
+                  Running with CUDA GPU acceleration. Switch back to CPU if needed (you can
+                  re-download later).
+                </p>
+                <Button onClick={handleSwitchToCpu} variant="outline" className="w-full" size="sm">
+                  <RotateCw className="h-4 w-4 mr-2" />
+                  Switch to CPU Backend
+                </Button>
+              </div>
+            )}
+            {error && (
+              <div className="flex items-center gap-2 text-sm text-destructive">
+                <AlertCircle className="h-4 w-4 shrink-0" />
+                <span>{error}</span>
+              </div>
+            )}
+          </>
+        )}
+
+        {/* CUDA download/manage section - show when no native GPU and not currently running CUDA */}
         {!hasNativeGpu && !isCurrentlyCuda && (
           <>
             {/* Download progress (manual download or auto-update) */}
@@ -315,7 +348,7 @@ export function GpuAcceleration() {
                 )}
 
                 {/* Downloaded but not active - show switch button */}
-                {cudaAvailable && !isCurrentlyCuda && platform.metadata.isTauri && (
+                {cudaAvailable && platform.metadata.isTauri && (
                   <div className="space-y-3">
                     <p className="text-sm text-muted-foreground">
                       CUDA backend is downloaded and ready. Restart the server to enable GPU
@@ -328,27 +361,8 @@ export function GpuAcceleration() {
                   </div>
                 )}
 
-                {/* Currently active - show switch back to CPU */}
-                {isCurrentlyCuda && platform.metadata.isTauri && (
-                  <div className="space-y-3">
-                    <p className="text-sm text-muted-foreground">
-                      Running with CUDA GPU acceleration. Switch back to CPU if needed (you can
-                      re-download later).
-                    </p>
-                    <Button
-                      onClick={handleSwitchToCpu}
-                      variant="outline"
-                      className="w-full"
-                      size="sm"
-                    >
-                      <RotateCw className="h-4 w-4 mr-2" />
-                      Switch to CPU Backend
-                    </Button>
-                  </div>
-                )}
-
                 {/* Delete option when downloaded (and not active) */}
-                {cudaAvailable && !isCurrentlyCuda && (
+                {cudaAvailable && (
                   <Button
                     onClick={handleDelete}
                     variant="ghost"

app/src/components/ServerSettings/ModelManagement.tsx

@@ -62,6 +62,10 @@ const MODEL_DESCRIPTIONS: Record<string, string> = {
     '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.',
+  'tada-1b':
+    'HumeAI TADA 1B — English speech-language model built on Llama 3.2 1B. Generates 700s+ of coherent audio with synchronized text-acoustic alignment.',
+  'tada-3b-ml':
+    'HumeAI TADA 3B Multilingual — built on Llama 3.2 3B. Supports 10 languages with high-fidelity voice cloning via text-acoustic dual alignment.',
   'whisper-base':
     'Smallest Whisper model (74M parameters). Fast transcription with moderate accuracy.',
   'whisper-small':
@@ -391,7 +395,8 @@ export function ModelManagement() {
       (m) =>
         m.model_name.startsWith('qwen-tts') ||
         m.model_name.startsWith('luxtts') ||
-        m.model_name.startsWith('chatterbox'),
+        m.model_name.startsWith('chatterbox') ||
+        m.model_name.startsWith('tada'),
     ) ?? [];
   const whisperModels = modelStatus?.models.filter((m) => m.model_name.startsWith('whisper')) ?? [];
 

app/src/lib/api/client.ts

@@ -32,8 +32,24 @@ import type {
   TranscriptionResponse,
   VoiceProfileCreate,
   VoiceProfileResponse,
+  WhisperModelSize,
 } from './types';
 
+function formatErrorDetail(detail: unknown, fallback: string): string {
+  if (typeof detail === 'string') return detail;
+  if (Array.isArray(detail)) {
+    return detail
+      .map((e: Record<string, unknown>) => e.msg || e.message || JSON.stringify(e))
+      .join('; ');
+  }
+  if (detail && typeof detail === 'object') {
+    const obj = detail as Record<string, unknown>;
+    if (typeof obj.message === 'string') return obj.message;
+    return JSON.stringify(detail);
+  }
+  return fallback;
+}
+
 class ApiClient {
   private getBaseUrl(): string {
     const serverUrl = useServerStore.getState().serverUrl;
@@ -54,7 +70,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.json();
@@ -113,7 +129,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.json();
@@ -147,7 +163,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.blob();
@@ -167,7 +183,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.json();
@@ -187,7 +203,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.json();
@@ -257,7 +273,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.blob();
@@ -271,7 +287,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.blob();
@@ -297,7 +313,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.json();
@@ -318,12 +334,19 @@ class ApiClient {
   }
 
   // Transcription
-  async transcribeAudio(file: File, language?: LanguageCode): Promise<TranscriptionResponse> {
+  async transcribeAudio(
+    file: File,
+    language?: LanguageCode,
+    model?: WhisperModelSize,
+  ): Promise<TranscriptionResponse> {
     const formData = new FormData();
     formData.append('file', file);
     if (language) {
       formData.append('language', language);
     }
+    if (model) {
+      formData.append('model', model);
+    }
 
     const url = `${this.getBaseUrl()}/transcribe`;
     const response = await fetch(url, {
@@ -335,7 +358,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.json();
@@ -608,7 +631,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.blob();
@@ -705,7 +728,7 @@ class ApiClient {
       const error = await response.json().catch(() => ({
         detail: response.statusText,
       }));
-      throw new Error(error.detail || `HTTP error! status: ${response.status}`);
+      throw new Error(formatErrorDetail(error.detail, `HTTP error! status: ${response.status}`));
     }
 
     return response.blob();

app/src/lib/api/types.ts

@@ -42,8 +42,8 @@ export interface GenerationRequest {
   text: string;
   language: LanguageCode;
   seed?: number;
-  model_size?: '1.7B' | '0.6B';
-  engine?: 'qwen' | 'luxtts' | 'chatterbox' | 'chatterbox_turbo';
+  model_size?: '1.7B' | '0.6B' | '1B' | '3B';
+  engine?: 'qwen' | 'luxtts' | 'chatterbox' | 'chatterbox_turbo' | 'tada';
   instruct?: string;
   max_chunk_chars?: number;
   crossfade_ms?: number;
@@ -99,8 +99,11 @@ export interface HistoryListResponse {
   total: number;
 }
 
+export type WhisperModelSize = 'base' | 'small' | 'medium' | 'large' | 'turbo';
+
 export interface TranscriptionRequest {
   language?: LanguageCode;
+  model?: WhisperModelSize;
 }
 
 export interface TranscriptionResponse {

app/src/lib/constants/languages.ts

@@ -66,6 +66,7 @@ export const ENGINE_LANGUAGES: Record<string, readonly LanguageCode[]> = {
     'zh',
   ],
   chatterbox_turbo: ['en'],
+  tada: ['en', 'ar', 'zh', 'de', 'es', 'fr', 'it', 'ja', 'pl', 'pt'],
 } as const;
 
 /** Helper: get language options for a given engine. */

app/src/lib/hooks/useGenerationForm.ts

@@ -15,9 +15,9 @@ const generationSchema = z.object({
   text: z.string().min(1, '').max(50000),
   language: z.enum(LANGUAGE_CODES as [LanguageCode, ...LanguageCode[]]),
   seed: z.number().int().optional(),
-  modelSize: z.enum(['1.7B', '0.6B']).optional(),
+  modelSize: z.enum(['1.7B', '0.6B', '1B', '3B']).optional(),
   instruct: z.string().max(500).optional(),
-  engine: z.enum(['qwen', 'luxtts', 'chatterbox', 'chatterbox_turbo']).optional(),
+  engine: z.enum(['qwen', 'luxtts', 'chatterbox', 'chatterbox_turbo', 'tada']).optional(),
 });
 
 export type GenerationFormValues = z.infer<typeof generationSchema>;
@@ -79,7 +79,11 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
             ? 'chatterbox-tts'
             : engine === 'chatterbox_turbo'
               ? 'chatterbox-turbo'
-              : `qwen-tts-${data.modelSize}`;
+              : engine === 'tada'
+                ? data.modelSize === '3B'
+                  ? 'tada-3b-ml'
+                  : 'tada-1b'
+                : `qwen-tts-${data.modelSize}`;
       const displayName =
         engine === 'luxtts'
           ? 'LuxTTS'
@@ -87,9 +91,13 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
             ? 'Chatterbox TTS'
             : engine === 'chatterbox_turbo'
               ? 'Chatterbox Turbo'
-              : data.modelSize === '1.7B'
-                ? 'Qwen TTS 1.7B'
-                : 'Qwen TTS 0.6B';
+              : engine === 'tada'
+                ? data.modelSize === '3B'
+                  ? 'TADA 3B Multilingual'
+                  : 'TADA 1B'
+                : data.modelSize === '1.7B'
+                  ? 'Qwen TTS 1.7B'
+                  : 'Qwen TTS 0.6B';
 
       // Check if model needs downloading
       try {
@@ -104,7 +112,7 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
         console.error('Failed to check model status:', error);
       }
 
-      const isQwen = engine === 'qwen';
+      const hasModelSizes = engine === 'qwen' || engine === 'tada';
       const effectsChain = options.getEffectsChain?.();
       // This now returns immediately with status="generating"
       const result = await generation.mutateAsync({
@@ -112,9 +120,9 @@ export function useGenerationForm(options: UseGenerationFormOptions = {}) {
         text: data.text,
         language: data.language,
         seed: data.seed,
-        model_size: isQwen ? data.modelSize : undefined,
+        model_size: hasModelSizes ? data.modelSize : undefined,
         engine,
-        instruct: isQwen ? data.instruct || undefined : undefined,
+        instruct: engine === 'qwen' ? data.instruct || undefined : undefined,
         max_chunk_chars: maxChunkChars,
         crossfade_ms: crossfadeMs,
         normalize: normalizeAudio,

app/src/lib/hooks/useGenerationProgress.ts

@@ -75,8 +75,8 @@ export function useGenerationProgress() {
             currentSources.delete(id);
             removePendingGeneration(id);
 
-            // Refresh history to pick up the completed generation
-            queryClient.invalidateQueries({ queryKey: ['history'] });
+            // Refetch history to pick up the completed generation
+            queryClient.refetchQueries({ queryKey: ['history'] });
 
             // If this generation was queued for a story, add it now
             const storyId = removePendingStoryAdd(id);
@@ -120,7 +120,7 @@ export function useGenerationProgress() {
             removePendingGeneration(id);
             removePendingStoryAdd(id);
 
-            queryClient.invalidateQueries({ queryKey: ['history'] });
+            queryClient.refetchQueries({ queryKey: ['history'] });
 
             toast({
               title: data.status === 'not_found' ? 'Generation not found' : 'Generation failed',
@@ -134,11 +134,12 @@ export function useGenerationProgress() {
       };
 
       source.onerror = () => {
-        // EventSource auto-reconnects, but if we get repeated errors
-        // just clean up
+        // SSE connection dropped — clean up and refresh history so any
+        // completed/failed generation still appears in the list
         source.close();
         currentSources.delete(id);
         removePendingGeneration(id);
+        queryClient.refetchQueries({ queryKey: ['history'] });
       };
 
       currentSources.set(id, source);

app/src/lib/hooks/useSystemAudioCapture.ts

@@ -1,4 +1,4 @@
-import { useState, useRef, useCallback, useEffect } from 'react';
+import { useCallback, useEffect, useRef, useState } from 'react';
 import { usePlatform } from '@/platform/PlatformContext';
 
 interface UseSystemAudioCaptureOptions {
@@ -94,15 +94,13 @@ export function useSystemAudioCapture({
       const blob = await platform.audio.stopSystemAudioCapture();
 
       // Pass the actual recorded duration
-      const recordedDuration = startTimeRef.current 
-        ? (Date.now() - startTimeRef.current) / 1000 
+      const recordedDuration = startTimeRef.current
+        ? (Date.now() - startTimeRef.current) / 1000
         : undefined;
       onRecordingComplete?.(blob, recordedDuration);
     } catch (err) {
       const errorMessage =
-        err instanceof Error
-          ? err.message
-          : 'Failed to stop system audio capture.';
+        err instanceof Error ? err.message : 'Failed to stop system audio capture.';
       setError(errorMessage);
     }
   }, [isRecording, onRecordingComplete, platform]);

app/src/lib/hooks/useTranscription.ts

@@ -1,10 +1,18 @@
 import { useMutation } from '@tanstack/react-query';
 import { apiClient } from '@/lib/api/client';
+import type { WhisperModelSize } from '@/lib/api/types';
 import type { LanguageCode } from '@/lib/constants/languages';
 
 export function useTranscription() {
   return useMutation({
-    mutationFn: ({ file, language }: { file: File; language?: LanguageCode }) =>
-      apiClient.transcribeAudio(file, language),
+    mutationFn: ({
+      file,
+      language,
+      model,
+    }: {
+      file: File;
+      language?: LanguageCode;
+      model?: WhisperModelSize;
+    }) => apiClient.transcribeAudio(file, language, model),
   });
 }

backend/__init__.py

@@ -1,3 +1,3 @@
 # Backend package
 
-__version__ = "0.2.3"
+__version__ = "0.3.1"

backend/backends/__init__.py

@@ -134,6 +134,7 @@ class STTBackend(Protocol):
         self,
         audio_path: str,
         language: Optional[str] = None,
+        model_size: Optional[str] = None,
     ) -> str:
         """
         Transcribe audio to text.
@@ -165,6 +166,7 @@ TTS_ENGINES = {
     "luxtts": "LuxTTS",
     "chatterbox": "Chatterbox TTS",
     "chatterbox_turbo": "Chatterbox Turbo",
+    "tada": "TADA",
 }
 
 
@@ -258,6 +260,24 @@ def _get_non_qwen_tts_configs() -> list[ModelConfig]:
             needs_trim=True,
             languages=["en"],
         ),
+        ModelConfig(
+            model_name="tada-1b",
+            display_name="TADA 1B (English)",
+            engine="tada",
+            hf_repo_id="HumeAI/tada-1b",
+            model_size="1B",
+            size_mb=4000,
+            languages=["en"],
+        ),
+        ModelConfig(
+            model_name="tada-3b-ml",
+            display_name="TADA 3B Multilingual",
+            engine="tada",
+            hf_repo_id="HumeAI/tada-3b-ml",
+            model_size="3B",
+            size_mb=8000,
+            languages=["en", "ar", "zh", "de", "es", "fr", "it", "ja", "pl", "pt"],
+        ),
     ]
 
 
@@ -338,10 +358,12 @@ def engine_has_model_sizes(engine: str) -> bool:
 
 
 async def load_engine_model(engine: str, model_size: str = "default") -> None:
-    """Load a model for the given engine, handling the Qwen model_size special case."""
+    """Load a model for the given engine, handling engines with multiple model sizes."""
     backend = get_tts_backend_for_engine(engine)
     if engine == "qwen":
         await backend.load_model_async(model_size)
+    elif engine == "tada":
+        await backend.load_model(model_size)
     else:
         await backend.load_model()
 
@@ -357,7 +379,7 @@ async def ensure_model_cached_or_raise(engine: str, model_size: str = "default")
             cfg = c
             break
 
-    if engine == "qwen":
+    if engine in ("qwen", "tada"):
         if not backend._is_model_cached(model_size):
             raise HTTPException(
                 status_code=400,
@@ -489,6 +511,10 @@ def get_tts_backend_for_engine(engine: str) -> TTSBackend:
             from .chatterbox_turbo_backend import ChatterboxTurboTTSBackend
 
             backend = ChatterboxTurboTTSBackend()
+        elif engine == "tada":
+            from .hume_backend import HumeTadaBackend
+
+            backend = HumeTadaBackend()
         else:
             raise ValueError(f"Unknown TTS engine: {engine}. Supported: {list(TTS_ENGINES.keys())}")
 

backend/backends/hume_backend.py

@@ -0,0 +1,347 @@
+"""
+HumeAI TADA TTS backend implementation.
+
+Wraps HumeAI's TADA (Text-Acoustic Dual Alignment) model for
+high-quality voice cloning. Two model variants:
+  - tada-1b: English-only, ~2B params (Llama 3.2 1B base)
+  - tada-3b-ml: Multilingual, ~4B params (Llama 3.2 3B base)
+
+Both use a shared encoder/codec (HumeAI/tada-codec). The encoder
+produces 1:1 aligned token embeddings from reference audio, and the
+causal LM generates speech via flow-matching diffusion.
+
+24kHz output, bf16 inference on CUDA, fp32 on CPU.
+"""
+
+import asyncio
+import logging
+import threading
+from typing import ClassVar, List, Optional, Tuple
+
+import numpy as np
+
+from . import TTSBackend
+from .base import (
+    is_model_cached,
+    get_torch_device,
+    combine_voice_prompts as _combine_voice_prompts,
+    model_load_progress,
+)
+from ..utils.cache import get_cache_key, get_cached_voice_prompt, cache_voice_prompt
+
+logger = logging.getLogger(__name__)
+
+# HuggingFace repos
+TADA_CODEC_REPO = "HumeAI/tada-codec"
+TADA_1B_REPO = "HumeAI/tada-1b"
+TADA_3B_ML_REPO = "HumeAI/tada-3b-ml"
+
+TADA_MODEL_REPOS = {
+    "1B": TADA_1B_REPO,
+    "3B": TADA_3B_ML_REPO,
+}
+
+# Key weight files for cache detection
+_TADA_MODEL_WEIGHT_FILES = [
+    "model.safetensors",
+]
+
+_TADA_CODEC_WEIGHT_FILES = [
+    "encoder/model.safetensors",
+]
+
+
+class HumeTadaBackend:
+    """HumeAI TADA TTS backend for high-quality voice cloning."""
+
+    _load_lock: ClassVar[threading.Lock] = threading.Lock()
+
+    def __init__(self):
+        self.model = None
+        self.encoder = None
+        self.model_size = "1B"  # default to 1B
+        self._device = None
+        self._model_load_lock = asyncio.Lock()
+
+    def _get_device(self) -> str:
+        # Force CPU on macOS — MPS has issues with flow matching
+        # and large vocab lm_head (>65536 output channels)
+        return get_torch_device(force_cpu_on_mac=True)
+
+    def is_loaded(self) -> bool:
+        return self.model is not None
+
+    def _get_model_path(self, model_size: str = "1B") -> str:
+        return TADA_MODEL_REPOS.get(model_size, TADA_1B_REPO)
+
+    def _is_model_cached(self, model_size: str = "1B") -> bool:
+        repo = TADA_MODEL_REPOS.get(model_size, TADA_1B_REPO)
+        model_cached = is_model_cached(repo, required_files=_TADA_MODEL_WEIGHT_FILES)
+        codec_cached = is_model_cached(TADA_CODEC_REPO, required_files=_TADA_CODEC_WEIGHT_FILES)
+        return model_cached and codec_cached
+
+    async def load_model(self, model_size: str = "1B") -> None:
+        """Load the TADA model and encoder."""
+        if self.model is not None and self.model_size == model_size:
+            return
+        async with self._model_load_lock:
+            if self.model is not None and self.model_size == model_size:
+                return
+            # Unload existing model if switching sizes
+            if self.model is not None:
+                self.unload_model()
+            self.model_size = model_size
+            await asyncio.to_thread(self._load_model_sync, model_size)
+
+    def _load_model_sync(self, model_size: str = "1B"):
+        """Synchronous model loading with progress tracking."""
+        model_name = f"tada-{model_size.lower()}"
+        is_cached = self._is_model_cached(model_size)
+        repo = TADA_MODEL_REPOS.get(model_size, TADA_1B_REPO)
+
+        with model_load_progress(model_name, is_cached):
+            # Install DAC shim before importing tada — tada's encoder/decoder
+            # import dac.nn.layers.Snake1d which requires the descript-audio-codec
+            # package.  The real package pulls in onnx/tensorboard/matplotlib via
+            # descript-audiotools, so we use a lightweight shim instead.
+            from ..utils.dac_shim import install_dac_shim
+            install_dac_shim()
+
+            import torch
+            from huggingface_hub import snapshot_download
+
+            device = self._get_device()
+            self._device = device
+            logger.info(f"Loading HumeAI TADA {model_size} on {device}...")
+
+            # Download codec (encoder + decoder) if not cached
+            logger.info("Downloading TADA codec...")
+            snapshot_download(
+                repo_id=TADA_CODEC_REPO,
+                token=None,
+                allow_patterns=["*.safetensors", "*.json", "*.txt", "*.bin"],
+            )
+
+            # Download model weights if not cached
+            logger.info(f"Downloading TADA {model_size} model...")
+            snapshot_download(
+                repo_id=repo,
+                token=None,
+                allow_patterns=["*.safetensors", "*.json", "*.txt", "*.bin", "*.model"],
+            )
+
+            # TADA hardcodes "meta-llama/Llama-3.2-1B" as the tokenizer
+            # source in its Aligner and TadaForCausalLM.from_pretrained().
+            # That repo is gated (requires Meta license acceptance).
+            # Download the tokenizer from an ungated mirror and get its
+            # local cache path so we can point TADA at it directly.
+            logger.info("Downloading Llama tokenizer (ungated mirror)...")
+            tokenizer_path = snapshot_download(
+                repo_id="unsloth/Llama-3.2-1B",
+                token=None,
+                allow_patterns=["tokenizer*", "special_tokens*"],
+            )
+
+            # Determine dtype — use bf16 on CUDA for ~50% memory savings
+            if device == "cuda" and torch.cuda.is_bf16_supported():
+                model_dtype = torch.bfloat16
+            else:
+                model_dtype = torch.float32
+
+            # Patch the Aligner config class to use the local tokenizer
+            # path instead of the gated "meta-llama/Llama-3.2-1B" default.
+            # This avoids monkey-patching AutoTokenizer.from_pretrained
+            # which corrupts the classmethod descriptor for other engines.
+            from tada.modules.aligner import AlignerConfig
+            AlignerConfig.tokenizer_name = tokenizer_path
+
+            # Load encoder (only needed for voice prompt encoding)
+            from tada.modules.encoder import Encoder
+            logger.info("Loading TADA encoder...")
+            self.encoder = Encoder.from_pretrained(
+                TADA_CODEC_REPO, subfolder="encoder"
+            ).to(device)
+            self.encoder.eval()
+
+            # Load the causal LM (includes decoder for wav generation).
+            # TadaForCausalLM.from_pretrained() calls
+            #   getattr(config, "tokenizer_name", "meta-llama/Llama-3.2-1B")
+            # which hits the gated repo. Pre-load the config from HF,
+            # inject the local tokenizer path, then pass it in.
+            from tada.modules.tada import TadaForCausalLM, TadaConfig
+            logger.info(f"Loading TADA {model_size} model...")
+            config = TadaConfig.from_pretrained(repo)
+            config.tokenizer_name = tokenizer_path
+            self.model = TadaForCausalLM.from_pretrained(
+                repo, config=config, torch_dtype=model_dtype
+            ).to(device)
+            self.model.eval()
+
+        logger.info(f"HumeAI TADA {model_size} loaded successfully on {device}")
+
+    def unload_model(self) -> None:
+        """Unload model and encoder to free memory."""
+        if self.model is not None:
+            del self.model
+            self.model = None
+        if self.encoder is not None:
+            del self.encoder
+            self.encoder = None
+
+        self._device = None
+
+        import torch
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+
+        logger.info("HumeAI TADA 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 using TADA's encoder.
+
+        TADA's encoder performs forced alignment between audio and text tokens,
+        producing an EncoderOutput with 1:1 token-audio alignment. If no
+        reference_text is provided, the encoder uses built-in ASR (English only).
+
+        We serialize the EncoderOutput to a dict for caching.
+        """
+        await self.load_model(self.model_size)
+
+        cache_key = (
+            "tada_" + get_cache_key(audio_path, reference_text)
+        ) if use_cache else None
+
+        if cache_key:
+            cached = get_cached_voice_prompt(cache_key)
+            if cached is not None and isinstance(cached, dict):
+                return cached, True
+
+        def _encode_sync():
+            import torch
+            import soundfile as sf
+
+            device = self._device
+
+            # Load audio with soundfile (torchaudio 2.10+ requires torchcodec)
+            audio_np, sr = sf.read(str(audio_path), dtype="float32")
+            audio = torch.from_numpy(audio_np).float()
+            if audio.ndim == 1:
+                audio = audio.unsqueeze(0)  # (samples,) -> (1, samples)
+            else:
+                audio = audio.T  # (samples, channels) -> (channels, samples)
+            audio = audio.to(device)
+
+            # Encode with forced alignment
+            text_arg = [reference_text] if reference_text else None
+            prompt = self.encoder(
+                audio, text=text_arg, sample_rate=sr
+            )
+
+            # Serialize EncoderOutput to a dict of CPU tensors for caching
+            prompt_dict = {}
+            for field_name in prompt.__dataclass_fields__:
+                val = getattr(prompt, field_name)
+                if isinstance(val, torch.Tensor):
+                    prompt_dict[field_name] = val.detach().cpu()
+                elif isinstance(val, list):
+                    prompt_dict[field_name] = val
+                elif isinstance(val, (int, float)):
+                    prompt_dict[field_name] = val
+                else:
+                    prompt_dict[field_name] = val
+            return prompt_dict
+
+        encoded = await asyncio.to_thread(_encode_sync)
+
+        if cache_key:
+            cache_voice_prompt(cache_key, encoded)
+
+        return encoded, False
+
+    async def combine_voice_prompts(
+        self,
+        audio_paths: List[str],
+        reference_texts: List[str],
+    ) -> Tuple[np.ndarray, str]:
+        return await _combine_voice_prompts(audio_paths, reference_texts, sample_rate=24000)
+
+    async def generate(
+        self,
+        text: str,
+        voice_prompt: dict,
+        language: str = "en",
+        seed: Optional[int] = None,
+        instruct: Optional[str] = None,
+    ) -> Tuple[np.ndarray, int]:
+        """
+        Generate audio from text using HumeAI TADA.
+
+        Args:
+            text: Text to synthesize
+            voice_prompt: Serialized EncoderOutput dict from create_voice_prompt()
+            language: Language code (en, ar, de, es, fr, it, ja, pl, pt, zh)
+            seed: Random seed for reproducibility
+            instruct: Not supported by TADA (ignored)
+
+        Returns:
+            Tuple of (audio_array, sample_rate=24000)
+        """
+        await self.load_model(self.model_size)
+
+        def _generate_sync():
+            import torch
+            from tada.modules.encoder import EncoderOutput
+
+            if seed is not None:
+                torch.manual_seed(seed)
+                if torch.cuda.is_available():
+                    torch.cuda.manual_seed(seed)
+
+            device = self._device
+
+            # Reconstruct EncoderOutput from the cached dict
+            restored = {}
+            for k, v in voice_prompt.items():
+                if isinstance(v, torch.Tensor):
+                    # Move to device and match model dtype for float tensors
+                    if v.is_floating_point():
+                        model_dtype = next(self.model.parameters()).dtype
+                        restored[k] = v.to(device=device, dtype=model_dtype)
+                    else:
+                        restored[k] = v.to(device=device)
+                else:
+                    restored[k] = v
+
+            prompt = EncoderOutput(**restored)
+
+            # For non-English with the 3B-ML model, we could reload the
+            # encoder with the language-specific aligner. However, the
+            # generation itself is language-agnostic — only the encoder's
+            # aligner changes. Since we encode at create_voice_prompt time,
+            # the language is already baked in. For simplicity, we don't
+            # reload the encoder here.
+
+            logger.info(f"[TADA] Generating ({language}), text length: {len(text)}")
+
+            output = self.model.generate(
+                prompt=prompt,
+                text=text,
+            )
+
+            # output.audio is a list of tensors (one per batch item)
+            if output.audio and output.audio[0] is not None:
+                audio_tensor = output.audio[0]
+                audio = audio_tensor.detach().cpu().numpy().squeeze().astype(np.float32)
+            else:
+                logger.warning("[TADA] Generation produced no audio")
+                audio = np.zeros(24000, dtype=np.float32)
+
+            return audio, 24000
+
+        return await asyncio.to_thread(_generate_sync)

backend/backends/mlx_backend.py

@@ -6,7 +6,6 @@ from typing import Optional, List, Tuple
 import asyncio
 import logging
 import numpy as np
-import os
 from pathlib import Path
 
 logger = logging.getLogger(__name__)
@@ -21,6 +20,7 @@ ensure_original_qwen_config_cached()
 from . import TTSBackend, STTBackend, LANGUAGE_CODE_TO_NAME, WHISPER_HF_REPOS
 from .base import is_model_cached, combine_voice_prompts as _combine_voice_prompts, model_load_progress
 from ..utils.cache import get_cache_key, get_cached_voice_prompt, cache_voice_prompt
+from ..utils.hf_offline_patch import force_offline_if_cached
 
 
 class MLXTTSBackend:
@@ -96,32 +96,13 @@ class MLXTTSBackend:
         model_name = f"qwen-tts-{model_size}"
         is_cached = self._is_model_cached(model_size)
 
-        # Force offline mode when cached to avoid network requests
-        original_hf_hub_offline = os.environ.get("HF_HUB_OFFLINE")
-        if is_cached:
-            os.environ["HF_HUB_OFFLINE"] = "1"
-            logger.info("[PATCH] Model %s is cached, forcing HF_HUB_OFFLINE=1 to avoid network requests", model_size)
+        with model_load_progress(model_name, is_cached):
+            from mlx_audio.tts import load
 
-        try:
-            with model_load_progress(model_name, is_cached):
-                from mlx_audio.tts import load
+            logger.info("Loading MLX TTS model %s...", model_size)
 
-                logger.info("Loading MLX TTS model %s...", model_size)
-
-                try:
-                    self.model = load(model_path)
-                except Exception as load_error:
-                    if is_cached and "offline" in str(load_error).lower():
-                        logger.warning("[PATCH] Offline load failed, trying with network: %s", load_error)
-                        os.environ.pop("HF_HUB_OFFLINE", None)
-                        self.model = load(model_path)
-                    else:
-                        raise
-        finally:
-            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)
+            with force_offline_if_cached(is_cached, model_name):
+                self.model = load(model_path)
 
         self._current_model_size = model_size
         self.model_size = model_size
@@ -329,7 +310,9 @@ class MLXSTTBackend:
 
             model_name = WHISPER_HF_REPOS.get(model_size, f"openai/whisper-{model_size}")
             logger.info("Loading MLX Whisper model %s...", model_size)
-            self.model = load(model_name)
+
+            with force_offline_if_cached(is_cached, progress_model_name):
+                self.model = load(model_name)
 
         self.model_size = model_size
         logger.info("MLX Whisper model %s loaded successfully", model_size)
@@ -345,18 +328,20 @@ class MLXSTTBackend:
         self,
         audio_path: str,
         language: Optional[str] = None,
+        model_size: Optional[str] = None,
     ) -> str:
         """
         Transcribe audio to text.
 
         Args:
             audio_path: Path to audio file
-            language: Optional language hint (en or zh)
+            language: Optional language hint
+            model_size: Optional model size override
 
         Returns:
             Transcribed text
         """
-        await self.load_model_async(None)
+        await self.load_model_async(model_size)
 
         def _transcribe_sync():
             """Run synchronous transcription in thread pool."""

backend/backends/pytorch_backend.py

@@ -19,6 +19,7 @@ from .base import (
 )
 from ..utils.cache import get_cache_key, get_cached_voice_prompt, cache_voice_prompt
 from ..utils.audio import load_audio
+from ..utils.hf_offline_patch import force_offline_if_cached
 
 
 class PyTorchTTSBackend:
@@ -96,18 +97,19 @@ class PyTorchTTSBackend:
             model_path = self._get_model_path(model_size)
             logger.info("Loading TTS model %s on %s...", model_size, self.device)
 
-            if self.device == "cpu":
-                self.model = Qwen3TTSModel.from_pretrained(
-                    model_path,
-                    torch_dtype=torch.float32,
-                    low_cpu_mem_usage=False,
-                )
-            else:
-                self.model = Qwen3TTSModel.from_pretrained(
-                    model_path,
-                    device_map=self.device,
-                    torch_dtype=torch.bfloat16,
-                )
+            with force_offline_if_cached(is_cached, model_name):
+                if self.device == "cpu":
+                    self.model = Qwen3TTSModel.from_pretrained(
+                        model_path,
+                        torch_dtype=torch.float32,
+                        low_cpu_mem_usage=False,
+                    )
+                else:
+                    self.model = Qwen3TTSModel.from_pretrained(
+                        model_path,
+                        device_map=self.device,
+                        torch_dtype=torch.bfloat16,
+                    )
 
         self._current_model_size = model_size
         self.model_size = model_size
@@ -282,8 +284,9 @@ class PyTorchSTTBackend:
             model_name = WHISPER_HF_REPOS.get(model_size, f"openai/whisper-{model_size}")
             logger.info("Loading Whisper model %s on %s...", model_size, self.device)
 
-            self.processor = WhisperProcessor.from_pretrained(model_name)
-            self.model = WhisperForConditionalGeneration.from_pretrained(model_name)
+            with force_offline_if_cached(is_cached, progress_model_name):
+                self.processor = WhisperProcessor.from_pretrained(model_name)
+                self.model = WhisperForConditionalGeneration.from_pretrained(model_name)
 
         self.model.to(self.device)
         self.model_size = model_size
@@ -306,18 +309,20 @@ class PyTorchSTTBackend:
         self,
         audio_path: str,
         language: Optional[str] = None,
+        model_size: Optional[str] = None,
     ) -> str:
         """
         Transcribe audio to text.
 
         Args:
             audio_path: Path to audio file
-            language: Optional language hint (en or zh)
+            language: Optional language hint
+            model_size: Optional model size override
 
         Returns:
             Transcribed text
         """
-        await self.load_model_async(None)
+        await self.load_model_async(model_size)
 
         def _transcribe_sync():
             """Run synchronous transcription in thread pool."""

backend/build_binary.py

@@ -34,9 +34,15 @@ def build_server(cuda=False):
     binary_name = "voicebox-server-cuda" if cuda else "voicebox-server"
 
     # PyInstaller arguments
+    # CUDA builds use --onedir so we can split the output into two archives:
+    #   1. Server core (~200-400MB) — versioned with the app
+    #   2. CUDA libs (~2GB) — versioned independently (only redownloaded on
+    #      CUDA toolkit / torch major version changes)
+    # CPU builds remain --onefile for simplicity.
+    pack_mode = "--onedir" if cuda else "--onefile"
     args = [
         "server.py",  # Use server.py as entry point instead of main.py
-        "--onefile",
+        pack_mode,
         "--name",
         binary_name,
     ]
@@ -165,9 +171,9 @@ def build_server(cuda=False):
             "tqdm",
             "--hidden-import",
             "requests",
-            "--collect-submodules",
-            "qwen_tts",
-            "--collect-data",
+            # qwen_tts uses inspect.getsource() at runtime to locate
+            # modeling_qwen3_tts.py — needs physical .py source files bundled
+            "--collect-all",
             "qwen_tts",
             # Fix for pkg_resources and jaraco namespace packages
             "--hidden-import",
@@ -186,6 +192,42 @@ def build_server(cuda=False):
             # needed by LuxTTS for text-to-phoneme conversion
             "--collect-all",
             "piper_phonemize",
+            # HumeAI TADA — speech-language model using Llama + flow matching
+            "--hidden-import",
+            "backend.backends.hume_backend",
+            "--hidden-import",
+            "tada",
+            "--hidden-import",
+            "tada.modules",
+            "--hidden-import",
+            "tada.modules.tada",
+            "--hidden-import",
+            "tada.modules.encoder",
+            "--hidden-import",
+            "tada.modules.decoder",
+            "--hidden-import",
+            "tada.modules.aligner",
+            "--hidden-import",
+            "tada.modules.acoustic_spkr_verf",
+            "--hidden-import",
+            "tada.nn",
+            "--hidden-import",
+            "tada.nn.vibevoice",
+            "--hidden-import",
+            "tada.utils",
+            "--hidden-import",
+            "tada.utils.gray_code",
+            "--hidden-import",
+            "tada.utils.text",
+            # DAC shim — provides dac.nn.layers.Snake1d without the real
+            # descript-audio-codec package (which pulls onnx/tensorboard via
+            # descript-audiotools). The shim is in backend/utils/dac_shim.py.
+            "--hidden-import",
+            "backend.utils.dac_shim",
+            "--hidden-import",
+            "torchaudio",
+            "--collect-submodules",
+            "tada",
         ]
     )
 
@@ -328,7 +370,7 @@ def build_server(cuda=False):
                     "torchvision",
                     "torchaudio",
                     "--index-url",
-                    "https://download.pytorch.org/whl/cu126",
+                    "https://download.pytorch.org/whl/cu128",
                     "--force-reinstall",
                     "-q",
                 ],

backend/models.py

@@ -66,9 +66,9 @@ class GenerationRequest(BaseModel):
     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|ar|da|el|fi|hi|ms|nl|no|pl|sv|sw|tr)$")
     seed: Optional[int] = Field(None, ge=0)
-    model_size: Optional[str] = Field(default="1.7B", pattern="^(1\\.7B|0\\.6B)$")
+    model_size: Optional[str] = Field(default="1.7B", pattern="^(1\\.7B|0\\.6B|1B|3B)$")
     instruct: Optional[str] = Field(None, max_length=500)
-    engine: Optional[str] = Field(default="qwen", pattern="^(qwen|luxtts|chatterbox|chatterbox_turbo)$")
+    engine: Optional[str] = Field(default="qwen", pattern="^(qwen|luxtts|chatterbox|chatterbox_turbo|tada)$")
     max_chunk_chars: int = Field(
         default=800, ge=100, le=5000, description="Max characters per chunk for long text splitting"
     )
@@ -149,7 +149,8 @@ class HistoryListResponse(BaseModel):
 class TranscriptionRequest(BaseModel):
     """Request model for audio transcription."""
 
-    language: Optional[str] = Field(None, pattern="^(en|zh)$")
+    language: Optional[str] = Field(None, pattern="^(en|zh|ja|ko|de|fr|ru|pt|es|it)$")
+    model: Optional[str] = Field(None, pattern="^(base|small|medium|large|turbo)$")
 
 
 class TranscriptionResponse(BaseModel):

backend/requirements.txt

@@ -8,7 +8,7 @@ sqlalchemy>=2.0.0
 alembic>=1.13.0
 
 # ML models
-torch>=2.1.0
+torch>=2.7.0
 transformers>=4.36.0,<=4.57.6
 accelerate>=0.26.0
 huggingface_hub>=0.20.0
@@ -33,6 +33,13 @@ s3tokenizer
 spacy-pkuseg
 pyloudnorm
 
+# HumeAI TADA sub-dependencies (hume-tada itself is installed
+# --no-deps in the setup script because it pins torch>=2.7,<2.8.
+# descript-audio-codec is NOT installed — it pulls onnx/tensorboard
+# via descript-audiotools.  A lightweight shim in utils/dac_shim.py
+# provides the only class TADA uses: Snake1d.)
+torchaudio
+
 # Audio processing
 librosa>=0.10.0
 soundfile>=0.12.0

backend/routes/generations.py

@@ -1,12 +1,15 @@
 """TTS generation endpoints."""
 
 import asyncio
+import logging
 import uuid
 
 from fastapi import APIRouter, Depends, HTTPException
 from fastapi.responses import StreamingResponse
 from sqlalchemy.orm import Session
 
+logger = logging.getLogger(__name__)
+
 from .. import models
 from ..services import history, profiles, tts
 from ..database import Generation as DBGeneration, VoiceProfile as DBVoiceProfile, get_db
@@ -181,25 +184,28 @@ async def get_generation_status(generation_id: str, db: Session = Depends(get_db
     import json
 
     async def event_stream():
-        while True:
-            db.expire_all()
-            gen = db.query(DBGeneration).filter_by(id=generation_id).first()
-            if not gen:
-                yield f"data: {json.dumps({'status': 'not_found', 'id': generation_id})}\n\n"
-                return
+        try:
+            while True:
+                db.expire_all()
+                gen = db.query(DBGeneration).filter_by(id=generation_id).first()
+                if not gen:
+                    yield f"data: {json.dumps({'status': 'not_found', 'id': generation_id})}\n\n"
+                    return
 
-            payload = {
-                "id": gen.id,
-                "status": gen.status or "completed",
-                "duration": gen.duration,
-                "error": gen.error,
-            }
-            yield f"data: {json.dumps(payload)}\n\n"
+                payload = {
+                    "id": gen.id,
+                    "status": gen.status or "completed",
+                    "duration": gen.duration,
+                    "error": gen.error,
+                }
+                yield f"data: {json.dumps(payload)}\n\n"
 
-            if (gen.status or "completed") in ("completed", "failed"):
-                return
+                if (gen.status or "completed") in ("completed", "failed"):
+                    return
 
-            await asyncio.sleep(1)
+                await asyncio.sleep(1)
+        except (BrokenPipeError, ConnectionResetError, asyncio.CancelledError):
+            logger.debug("SSE client disconnected for generation %s", generation_id)
 
     return StreamingResponse(
         event_stream(),
@@ -265,9 +271,12 @@ async def stream_speech(
     wav_bytes = tts.audio_to_wav_bytes(audio, sample_rate)
 
     async def _wav_stream():
-        chunk_size = 64 * 1024
-        for i in range(0, len(wav_bytes), chunk_size):
-            yield wav_bytes[i : i + chunk_size]
+        try:
+            chunk_size = 64 * 1024
+            for i in range(0, len(wav_bytes), chunk_size):
+                yield wav_bytes[i : i + chunk_size]
+        except (BrokenPipeError, ConnectionResetError, asyncio.CancelledError):
+            logger.debug("Client disconnected during audio stream")
 
     return StreamingResponse(
         _wav_stream(),

backend/routes/profiles.py

@@ -102,6 +102,10 @@ async def delete_profile(
     return {"message": "Profile deleted successfully"}
 
 
+SAMPLE_MAX_FILE_SIZE = 50 * 1024 * 1024  # 50 MB
+SAMPLE_UPLOAD_CHUNK_SIZE = 1024 * 1024  # 1 MB
+
+
 @router.post("/profiles/{profile_id}/samples", response_model=models.ProfileSampleResponse)
 async def add_profile_sample(
     profile_id: str,
@@ -115,8 +119,16 @@ async def add_profile_sample(
     file_suffix = _uploaded_ext if _uploaded_ext in _allowed_audio_exts else ".wav"
 
     with tempfile.NamedTemporaryFile(suffix=file_suffix, delete=False) as tmp:
-        content = await file.read()
-        tmp.write(content)
+        total_size = 0
+        while chunk := await file.read(SAMPLE_UPLOAD_CHUNK_SIZE):
+            total_size += len(chunk)
+            if total_size > SAMPLE_MAX_FILE_SIZE:
+                Path(tmp.name).unlink(missing_ok=True)
+                raise HTTPException(
+                    status_code=413,
+                    detail=f"File too large (max {SAMPLE_MAX_FILE_SIZE // (1024 * 1024)} MB)",
+                )
+            tmp.write(chunk)
         tmp_path = tmp.name
 
     try:

backend/routes/transcription.py

@@ -20,6 +20,7 @@ UPLOAD_CHUNK_SIZE = 1024 * 1024  # 1MB
 async def transcribe_audio(
     file: UploadFile = File(...),
     language: str | None = Form(None),
+    model: str | None = Form(None),
 ):
     """Transcribe audio file to text."""
     with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
@@ -29,14 +30,23 @@ async def transcribe_audio(
 
     try:
         from ..utils.audio import load_audio
+        from ..backends import WHISPER_HF_REPOS
 
         audio, sr = await asyncio.to_thread(load_audio, tmp_path)
         duration = len(audio) / sr
 
         whisper_model = transcribe.get_whisper_model()
-        model_size = whisper_model.model_size
+        model_size = model if model else whisper_model.model_size
 
-        if not whisper_model.is_loaded() and not whisper_model._is_model_cached(model_size):
+        valid_sizes = list(WHISPER_HF_REPOS.keys())
+        if model_size not in valid_sizes:
+            raise HTTPException(
+                status_code=400,
+                detail=f"Invalid model size '{model_size}'. Must be one of: {', '.join(valid_sizes)}",
+            )
+
+        already_loaded = whisper_model.is_loaded() and whisper_model.model_size == model_size
+        if not already_loaded and not whisper_model._is_model_cached(model_size):
             progress_model_name = f"whisper-{model_size}"
             task_manager = get_task_manager()
 
@@ -59,7 +69,7 @@ async def transcribe_audio(
                 },
             )
 
-        text = await whisper_model.transcribe(tmp_path, language)
+        text = await whisper_model.transcribe(tmp_path, language, model_size)
 
         return models.TranscriptionResponse(
             text=text,

backend/services/cuda.py

@@ -1,16 +1,22 @@
 """
-CUDA backend binary download, assembly, and verification.
+CUDA backend download, assembly, and verification.
 
-Downloads split parts of the CUDA-enabled voicebox-server binary from
-GitHub Releases, reassembles them, verifies integrity via SHA-256,
-and places the binary in the app's data directory for use on next
-backend restart.
+Downloads two archives from GitHub Releases:
+  1. Server core (voicebox-server-cuda.tar.gz) — the exe + non-NVIDIA deps,
+     versioned with the app.
+  2. CUDA libs (cuda-libs-{version}.tar.gz) — NVIDIA runtime libraries,
+     versioned independently (only redownloaded on CUDA toolkit bump).
+
+Both archives are extracted into {data_dir}/backends/cuda/ which forms the
+complete PyInstaller --onedir directory structure that torch expects.
 """
 
 import hashlib
+import json
 import logging
 import os
 import sys
+import tarfile
 from pathlib import Path
 from typing import Optional
 
@@ -24,6 +30,10 @@ GITHUB_RELEASES_URL = "https://github.com/jamiepine/voicebox/releases/download"
 
 PROGRESS_KEY = "cuda-backend"
 
+# The current expected CUDA libs version.  Bump this when we change the
+# CUDA toolkit version or torch's CUDA dependency changes (e.g. cu126 -> cu128).
+CUDA_LIBS_VERSION = "cu128-v1"
+
 
 def get_backends_dir() -> Path:
     """Directory where downloaded backend binaries are stored."""
@@ -32,21 +42,46 @@ def get_backends_dir() -> Path:
     return d
 
 
-def get_cuda_binary_name() -> str:
-    """Platform-specific CUDA binary filename."""
+def get_cuda_dir() -> Path:
+    """Directory where the CUDA backend (onedir) is extracted."""
+    d = get_backends_dir() / "cuda"
+    d.mkdir(parents=True, exist_ok=True)
+    return d
+
+
+def get_cuda_exe_name() -> str:
+    """Platform-specific CUDA executable filename."""
     if sys.platform == "win32":
         return "voicebox-server-cuda.exe"
     return "voicebox-server-cuda"
 
 
 def get_cuda_binary_path() -> Optional[Path]:
-    """Return path to CUDA binary if it exists."""
-    p = get_backends_dir() / get_cuda_binary_name()
+    """Return path to the CUDA executable if it exists inside the onedir."""
+    p = get_cuda_dir() / get_cuda_exe_name()
     if p.exists():
         return p
     return None
 
 
+def get_cuda_libs_manifest_path() -> Path:
+    """Path to the cuda-libs.json manifest inside the CUDA dir."""
+    return get_cuda_dir() / "cuda-libs.json"
+
+
+def get_installed_cuda_libs_version() -> Optional[str]:
+    """Read the installed CUDA libs version from cuda-libs.json, or None."""
+    manifest_path = get_cuda_libs_manifest_path()
+    if not manifest_path.exists():
+        return None
+    try:
+        data = json.loads(manifest_path.read_text())
+        return data.get("version")
+    except Exception as e:
+        logger.warning(f"Could not read cuda-libs.json: {e}")
+        return None
+
+
 def is_cuda_active() -> bool:
     """Check if the current process is the CUDA binary.
 
@@ -60,25 +95,151 @@ def get_cuda_status() -> dict:
     progress_manager = get_progress_manager()
     cuda_path = get_cuda_binary_path()
     progress = progress_manager.get_progress(PROGRESS_KEY)
+    cuda_libs_version = get_installed_cuda_libs_version()
 
     return {
         "available": cuda_path is not None,
         "active": is_cuda_active(),
         "binary_path": str(cuda_path) if cuda_path else None,
+        "cuda_libs_version": cuda_libs_version,
         "downloading": progress is not None and progress.get("status") == "downloading",
         "download_progress": progress,
     }
 
 
-async def download_cuda_binary(version: Optional[str] = None):
-    """Download the CUDA backend binary from GitHub Releases.
+def _needs_server_download(version: Optional[str] = None) -> bool:
+    """Check if the server core archive needs to be (re)downloaded."""
+    cuda_path = get_cuda_binary_path()
+    if not cuda_path:
+        return True
+    # Check if the binary version matches the expected app version
+    installed = get_cuda_binary_version()
+    expected = version or __version__
+    if expected.startswith("v"):
+        expected = expected[1:]
+    return installed != expected
 
-    Downloads split parts listed in a manifest file, concatenates them,
-    and verifies the SHA-256 checksum for integrity. Atomic write
-    (temp file -> rename).
+
+def _needs_cuda_libs_download() -> bool:
+    """Check if the CUDA libs archive needs to be (re)downloaded."""
+    installed = get_installed_cuda_libs_version()
+    if installed is None:
+        return True
+    return installed != CUDA_LIBS_VERSION
+
+
+async def _download_and_extract_archive(
+    client,
+    url: str,
+    sha256_url: Optional[str],
+    dest_dir: Path,
+    label: str,
+    progress_offset: int,
+    total_size: int,
+):
+    """Download a .tar.gz archive and extract it into dest_dir.
 
     Args:
-        version: Version tag (e.g. "v0.2.0"). Defaults to current app version.
+        client: httpx.AsyncClient
+        url: URL of the .tar.gz archive
+        sha256_url: URL of the .sha256 checksum file (optional)
+        dest_dir: Directory to extract into
+        label: Human-readable label for progress updates
+        progress_offset: Byte offset for progress reporting (when downloading
+            multiple archives sequentially)
+        total_size: Total bytes across all downloads (for progress bar)
+    """
+    progress = get_progress_manager()
+    temp_path = dest_dir / f".download-{label.replace(' ', '-')}.tmp"
+
+    # Clean up leftover partial download
+    if temp_path.exists():
+        temp_path.unlink()
+
+    # Fetch expected checksum (fail-fast: never extract an unverified archive)
+    expected_sha = None
+    if sha256_url:
+        try:
+            sha_resp = await client.get(sha256_url)
+            sha_resp.raise_for_status()
+            expected_sha = sha_resp.text.strip().split()[0]
+            logger.info(f"{label}: expected SHA-256: {expected_sha[:16]}...")
+        except Exception as e:
+            raise RuntimeError(f"{label}: failed to fetch checksum from {sha256_url}") from e
+
+    # Stream download, verify, and extract — always clean up temp file
+    downloaded = 0
+    try:
+        async with client.stream("GET", url) as response:
+            response.raise_for_status()
+            with open(temp_path, "wb") as f:
+                async for chunk in response.aiter_bytes(chunk_size=1024 * 1024):
+                    f.write(chunk)
+                    downloaded += len(chunk)
+                    progress.update_progress(
+                        PROGRESS_KEY,
+                        current=progress_offset + downloaded,
+                        total=total_size,
+                        filename=f"Downloading {label}",
+                        status="downloading",
+                    )
+
+        # Verify integrity
+        if expected_sha:
+            progress.update_progress(
+                PROGRESS_KEY,
+                current=progress_offset + downloaded,
+                total=total_size,
+                filename=f"Verifying {label}...",
+                status="downloading",
+            )
+            sha256 = hashlib.sha256()
+            with open(temp_path, "rb") as f:
+                while True:
+                    data = f.read(1024 * 1024)
+                    if not data:
+                        break
+                    sha256.update(data)
+            actual = sha256.hexdigest()
+            if actual != expected_sha:
+                raise ValueError(
+                    f"{label} integrity check failed: expected {expected_sha[:16]}..., got {actual[:16]}..."
+                )
+            logger.info(f"{label}: integrity verified")
+
+        # Extract (use data filter for path traversal protection on Python 3.12+)
+        progress.update_progress(
+            PROGRESS_KEY,
+            current=progress_offset + downloaded,
+            total=total_size,
+            filename=f"Extracting {label}...",
+            status="downloading",
+        )
+        with tarfile.open(temp_path, "r:gz") as tar:
+            if sys.version_info >= (3, 12):
+                tar.extractall(path=dest_dir, filter="data")
+            else:
+                tar.extractall(path=dest_dir)
+
+        logger.info(f"{label}: extracted to {dest_dir}")
+    finally:
+        if temp_path.exists():
+            temp_path.unlink()
+    return downloaded
+
+
+async def download_cuda_binary(version: Optional[str] = None):
+    """Download the CUDA backend (server core + CUDA libs if needed).
+
+    Downloads both archives from GitHub Releases, extracts them into
+    {data_dir}/backends/cuda/, and writes the cuda-libs.json manifest.
+
+    Only downloads what's needed:
+    - Server core: always redownloaded (versioned with app)
+    - CUDA libs: only if missing or version mismatch
+
+    Args:
+        version: Version tag (e.g. "v0.3.0"). Defaults to current app version.
     """
     import httpx
 
@@ -86,114 +247,91 @@ async def download_cuda_binary(version: Optional[str] = None):
         version = f"v{__version__}"
 
     progress = get_progress_manager()
-    binary_name = get_cuda_binary_name()
-    dest_dir = get_backends_dir()
-    final_path = dest_dir / binary_name
-    temp_path = dest_dir / f"{binary_name}.download"
+    cuda_dir = get_cuda_dir()
 
-    # Clean up any leftover partial download
-    if temp_path.exists():
-        temp_path.unlink()
+    need_server = _needs_server_download(version)
+    need_libs = _needs_cuda_libs_download()
 
-    logger.info(f"Starting CUDA backend download for {version}")
+    if not need_server and not need_libs:
+        logger.info("CUDA backend is up to date, nothing to download")
+        return
+
+    logger.info(
+        f"Starting CUDA backend download for {version} "
+        f"(server={'yes' if need_server else 'cached'}, "
+        f"libs={'yes' if need_libs else 'cached'})"
+    )
     progress.update_progress(
-        PROGRESS_KEY, current=0, total=0,
-        filename="Fetching manifest...", status="downloading",
+        PROGRESS_KEY,
+        current=0,
+        total=0,
+        filename="Preparing download...",
+        status="downloading",
     )
 
     base_url = f"{GITHUB_RELEASES_URL}/{version}"
-    stem = Path(binary_name).stem  # voicebox-server-cuda
+    server_archive = "voicebox-server-cuda.tar.gz"
+    libs_archive = f"cuda-libs-{CUDA_LIBS_VERSION}.tar.gz"
 
     try:
         async with httpx.AsyncClient(follow_redirects=True, timeout=30.0) as client:
-            # Fetch the manifest (list of split part filenames)
-            manifest_url = f"{base_url}/{stem}.manifest"
-            manifest_resp = await client.get(manifest_url)
-            manifest_resp.raise_for_status()
-            parts = [p.strip() for p in manifest_resp.text.strip().splitlines() if p.strip()]
-
-            if not parts:
-                raise ValueError("Empty manifest — no split parts found")
-
-            logger.info(f"Found {len(parts)} split parts to download")
-
-            # Fetch expected checksum (optional — for integrity verification)
-            expected_sha = None
-            try:
-                sha_url = f"{base_url}/{stem}.sha256"
-                sha_resp = await client.get(sha_url)
-                if sha_resp.status_code == 200:
-                    # Format: "sha256hex  filename\n"
-                    expected_sha = sha_resp.text.strip().split()[0]
-                    logger.info(f"Expected SHA-256: {expected_sha[:16]}...")
-            except Exception as e:
-                logger.warning(f"Could not fetch checksum file — skipping verification: {e}")
-
-            # Get total size across all parts by issuing HEAD requests
+            # Estimate total download size
             total_size = 0
-            for part_name in parts:
+            if need_server:
                 try:
-                    head_resp = await client.head(f"{base_url}/{part_name}")
-                    content_length = int(head_resp.headers.get("content-length", 0))
-                    total_size += content_length
+                    head = await client.head(f"{base_url}/{server_archive}")
+                    total_size += int(head.headers.get("content-length", 0))
                 except Exception:
                     pass
+            if need_libs:
+                try:
+                    head = await client.head(f"{base_url}/{libs_archive}")
+                    total_size += int(head.headers.get("content-length", 0))
+                except Exception:
+                    pass
+
             logger.info(f"Total download size: {total_size / 1024 / 1024:.1f} MB")
 
-            # Download and concatenate parts
-            total_downloaded = 0
-            with open(temp_path, "wb") as f:
-                for i, part_name in enumerate(parts):
-                    part_url = f"{base_url}/{part_name}"
-                    logger.info(f"Downloading part {i + 1}/{len(parts)}: {part_name}")
+            offset = 0
 
-                    async with client.stream("GET", part_url) as response:
-                        response.raise_for_status()
-                        async for chunk in response.aiter_bytes(chunk_size=1024 * 1024):
-                            f.write(chunk)
-                            total_downloaded += len(chunk)
-                            progress.update_progress(
-                                PROGRESS_KEY, current=total_downloaded, total=total_size,
-                                filename=f"Downloading CUDA backend ({i + 1}/{len(parts)})",
-                                status="downloading",
-                            )
-
-        # Verify integrity if checksum was available
-        if expected_sha:
-            progress.update_progress(
-                PROGRESS_KEY, current=total_downloaded, total=total_downloaded,
-                filename="Verifying integrity...", status="downloading",
-            )
-            sha256 = hashlib.sha256()
-            with open(temp_path, "rb") as f:
-                while True:
-                    chunk = f.read(1024 * 1024)
-                    if not chunk:
-                        break
-                    sha256.update(chunk)
-
-            actual = sha256.hexdigest()
-            if actual != expected_sha:
-                raise ValueError(
-                    f"Integrity check failed: expected {expected_sha[:16]}..., "
-                    f"got {actual[:16]}..."
+            # Download server core
+            if need_server:
+                server_downloaded = await _download_and_extract_archive(
+                    client,
+                    url=f"{base_url}/{server_archive}",
+                    sha256_url=f"{base_url}/{server_archive}.sha256",
+                    dest_dir=cuda_dir,
+                    label="CUDA server",
+                    progress_offset=offset,
+                    total_size=total_size,
                 )
-            logger.info(f"Integrity verified: {actual[:16]}...")
+                offset += server_downloaded
 
-        # Atomic move into place (replace handles existing target on all platforms)
-        temp_path.replace(final_path)
+                # Make executable on Unix
+                exe_path = cuda_dir / get_cuda_exe_name()
+                if sys.platform != "win32" and exe_path.exists():
+                    exe_path.chmod(0o755)
 
-        # Make executable on Unix
-        if sys.platform != "win32":
-            final_path.chmod(0o755)
+            # Download CUDA libs
+            if need_libs:
+                await _download_and_extract_archive(
+                    client,
+                    url=f"{base_url}/{libs_archive}",
+                    sha256_url=f"{base_url}/{libs_archive}.sha256",
+                    dest_dir=cuda_dir,
+                    label="CUDA libraries",
+                    progress_offset=offset,
+                    total_size=total_size,
+                )
 
-        logger.info(f"CUDA backend downloaded to {final_path}")
+                # Write local cuda-libs.json manifest
+                manifest = {"version": CUDA_LIBS_VERSION}
+                get_cuda_libs_manifest_path().write_text(json.dumps(manifest, indent=2) + "\n")
+
+        logger.info(f"CUDA backend ready at {cuda_dir}")
         progress.mark_complete(PROGRESS_KEY)
 
     except Exception as e:
-        # Clean up on failure
-        if temp_path.exists():
-            temp_path.unlink()
         logger.error(f"CUDA backend download failed: {e}")
         progress.mark_error(PROGRESS_KEY, str(e))
         raise
@@ -202,15 +340,19 @@ async def download_cuda_binary(version: Optional[str] = None):
 def get_cuda_binary_version() -> Optional[str]:
     """Get the version of the installed CUDA binary, or None if not installed."""
     import subprocess
+
     cuda_path = get_cuda_binary_path()
     if not cuda_path:
         return None
     try:
         result = subprocess.run(
             [str(cuda_path), "--version"],
-            capture_output=True, text=True, timeout=30,
+            capture_output=True,
+            text=True,
+            timeout=30,
+            cwd=str(cuda_path.parent),  # Run from the onedir directory
         )
-        # Output format: "voicebox-server 0.2.0"
+        # Output format: "voicebox-server 0.3.0"
         for line in result.stdout.strip().splitlines():
             if "voicebox-server" in line:
                 return line.split()[-1]
@@ -222,26 +364,29 @@ def get_cuda_binary_version() -> Optional[str]:
 async def check_and_update_cuda_binary():
     """Check if the CUDA binary is outdated and auto-download if so.
 
-    Called on server startup. If a CUDA binary exists but its version
-    doesn't match the current app version, triggers a background download
-    of the updated CUDA binary. The download progress is visible to the
-    frontend via the existing SSE progress endpoint.
+    Called on server startup. Checks both server version and CUDA libs
+    version. Downloads only what's needed.
     """
     cuda_path = get_cuda_binary_path()
     if not cuda_path:
         return  # No CUDA binary installed, nothing to update
 
-    cuda_version = get_cuda_binary_version()
-    current_version = __version__
+    need_server = _needs_server_download()
+    need_libs = _needs_cuda_libs_download()
 
-    if cuda_version == current_version:
-        logger.info(f"CUDA binary is up to date (v{current_version})")
+    if not need_server and not need_libs:
+        logger.info(f"CUDA binary is up to date (server=v{__version__}, libs={get_installed_cuda_libs_version()})")
         return
 
-    logger.info(
-        f"CUDA binary version mismatch: binary=v{cuda_version}, app=v{current_version}. "
-        f"Auto-downloading updated CUDA backend..."
-    )
+    reasons = []
+    if need_server:
+        cuda_version = get_cuda_binary_version()
+        reasons.append(f"server v{cuda_version} != v{__version__}")
+    if need_libs:
+        installed_libs = get_installed_cuda_libs_version()
+        reasons.append(f"libs {installed_libs} != {CUDA_LIBS_VERSION}")
+
+    logger.info(f"CUDA backend needs update ({', '.join(reasons)}). Auto-downloading...")
 
     try:
         await download_cuda_binary()
@@ -250,10 +395,12 @@ async def check_and_update_cuda_binary():
 
 
 async def delete_cuda_binary() -> bool:
-    """Delete the downloaded CUDA binary. Returns True if deleted."""
-    path = get_cuda_binary_path()
-    if path and path.exists():
-        path.unlink()
-        logger.info(f"Deleted CUDA binary: {path}")
+    """Delete the downloaded CUDA backend directory. Returns True if deleted."""
+    import shutil
+
+    cuda_dir = get_cuda_dir()
+    if cuda_dir.exists() and any(cuda_dir.iterdir()):
+        shutil.rmtree(cuda_dir)
+        logger.info(f"Deleted CUDA backend directory: {cuda_dir}")
         return True
     return False

backend/services/profiles.py

@@ -22,7 +22,7 @@ from ..database import (
     Generation as DBGeneration,
 )
 from ..models import EffectConfig
-from ..utils.audio import validate_reference_audio, load_audio, save_audio
+from ..utils.audio import validate_reference_audio, validate_and_load_reference_audio, load_audio, save_audio
 from ..utils.images import validate_image, process_avatar
 from ..utils.cache import _get_cache_dir, clear_profile_cache
 from .tts import get_tts_model
@@ -117,11 +117,16 @@ async def add_profile_sample(
     Returns:
         Created sample
     """
+    import asyncio
+
     profile = db.query(DBVoiceProfile).filter_by(id=profile_id).first()
     if not profile:
         raise ValueError(f"Profile {profile_id} not found")
 
-    is_valid, error_msg = validate_reference_audio(audio_path)
+    # Validate and load audio in a single pass, off the event loop
+    is_valid, error_msg, audio, sr = await asyncio.to_thread(
+        validate_and_load_reference_audio, audio_path
+    )
     if not is_valid:
         raise ValueError(f"Invalid reference audio: {error_msg}")
 
@@ -130,8 +135,7 @@ async def add_profile_sample(
     profile_dir.mkdir(parents=True, exist_ok=True)
 
     dest_path = profile_dir / f"{sample_id}.wav"
-    audio, sr = load_audio(audio_path)
-    save_audio(audio, str(dest_path), sr)
+    await asyncio.to_thread(save_audio, audio, str(dest_path), sr)
 
     db_sample = DBProfileSample(
         id=sample_id,

backend/utils/audio.py

@@ -217,22 +217,40 @@ def validate_reference_audio(
     Returns:
         Tuple of (is_valid, error_message)
     """
+    result = validate_and_load_reference_audio(
+        audio_path, min_duration, max_duration, min_rms
+    )
+    return (result[0], result[1])
+
+
+def validate_and_load_reference_audio(
+    audio_path: str,
+    min_duration: float = 2.0,
+    max_duration: float = 30.0,
+    min_rms: float = 0.01,
+) -> Tuple[bool, Optional[str], Optional[np.ndarray], Optional[int]]:
+    """
+    Validate and load reference audio in a single pass.
+    
+    Returns:
+        Tuple of (is_valid, error_message, audio_array, sample_rate)
+    """
     try:
         audio, sr = load_audio(audio_path)
         duration = len(audio) / sr
         
         if duration < min_duration:
-            return False, f"Audio too short (minimum {min_duration} seconds)"
+            return False, f"Audio too short (minimum {min_duration} seconds)", None, None
         if duration > max_duration:
-            return False, f"Audio too long (maximum {max_duration} seconds)"
+            return False, f"Audio too long (maximum {max_duration} seconds)", None, None
         
         rms = np.sqrt(np.mean(audio**2))
         if rms < min_rms:
-            return False, "Audio is too quiet or silent"
+            return False, "Audio is too quiet or silent", None, None
         
         if np.abs(audio).max() > 0.99:
-            return False, "Audio is clipping (reduce input gain)"
+            return False, "Audio is clipping (reduce input gain)", None, None
         
-        return True, None
+        return True, None, audio, sr
     except Exception as e:
-        return False, f"Error validating audio: {str(e)}"
+        return False, f"Error validating audio: {str(e)}", None, None

backend/utils/dac_shim.py

@@ -0,0 +1,95 @@
+"""
+Minimal shim for descript-audio-codec (DAC).
+
+TADA only imports Snake1d from dac.nn.layers and dac.model.dac.
+The real DAC package pulls in descript-audiotools which depends on
+onnx, tensorboard, protobuf, matplotlib, pystoi, etc. — none of
+which are needed for TADA's runtime use of Snake1d.
+
+This shim provides the exact Snake1d implementation (MIT-licensed,
+from https://github.com/descriptinc/descript-audio-codec) so we can
+avoid the entire audiotools dependency chain.
+
+If the real DAC package is installed, this module is never used —
+Python's import system will find the site-packages version first.
+Install this shim only when descript-audio-codec is NOT installed.
+"""
+
+import sys
+import types
+
+import torch
+import torch.nn as nn
+
+
+# ── Snake activation (from dac/nn/layers.py) ────────────────────────
+
+# NOTE: The original DAC code uses @torch.jit.script here for a 1.4x
+# speedup.  We omit it because TorchScript calls inspect.getsource()
+# which fails inside a PyInstaller frozen binary (no .py source files).
+def snake(x: torch.Tensor, alpha: torch.Tensor) -> torch.Tensor:
+    shape = x.shape
+    x = x.reshape(shape[0], shape[1], -1)
+    x = x + (alpha + 1e-9).reciprocal() * torch.sin(alpha * x).pow(2)
+    x = x.reshape(shape)
+    return x
+
+
+class Snake1d(nn.Module):
+    def __init__(self, channels: int):
+        super().__init__()
+        self.alpha = nn.Parameter(torch.ones(1, channels, 1))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return snake(x, self.alpha)
+
+
+# ── Register as dac.nn.layers and dac.model.dac ─────────────────────
+
+def install_dac_shim() -> None:
+    """Register fake dac package modules in sys.modules.
+
+    Only installs the shim if 'dac' is not already importable
+    (i.e. the real descript-audio-codec is not installed).
+    """
+    try:
+        import dac  # noqa: F401  — real package exists, do nothing
+        return
+    except ImportError:
+        pass
+
+    # Create the module tree: dac -> dac.nn -> dac.nn.layers
+    #                              -> dac.model -> dac.model.dac
+    dac_pkg = types.ModuleType("dac")
+    dac_pkg.__path__ = []  # make it a package
+    dac_pkg.__package__ = "dac"
+
+    dac_nn = types.ModuleType("dac.nn")
+    dac_nn.__path__ = []
+    dac_nn.__package__ = "dac.nn"
+
+    dac_nn_layers = types.ModuleType("dac.nn.layers")
+    dac_nn_layers.__package__ = "dac.nn"
+    dac_nn_layers.Snake1d = Snake1d
+    dac_nn_layers.snake = snake
+
+    dac_model = types.ModuleType("dac.model")
+    dac_model.__path__ = []
+    dac_model.__package__ = "dac.model"
+
+    dac_model_dac = types.ModuleType("dac.model.dac")
+    dac_model_dac.__package__ = "dac.model"
+    dac_model_dac.Snake1d = Snake1d
+
+    # Wire up submodules
+    dac_pkg.nn = dac_nn
+    dac_pkg.model = dac_model
+    dac_nn.layers = dac_nn_layers
+    dac_model.dac = dac_model_dac
+
+    # Register in sys.modules
+    sys.modules["dac"] = dac_pkg
+    sys.modules["dac.nn"] = dac_nn
+    sys.modules["dac.nn.layers"] = dac_nn_layers
+    sys.modules["dac.model"] = dac_model
+    sys.modules["dac.model.dac"] = dac_model_dac

backend/utils/hf_offline_patch.py

@@ -1,17 +1,64 @@
 """Monkey-patch huggingface_hub to force offline mode with cached models.
 
-Prevents mlx_audio from making network requests when models are already
-downloaded. Must be imported BEFORE mlx_audio.
+Prevents mlx_audio / transformers from making network requests when models
+are already downloaded. Must be imported BEFORE mlx_audio.
 """
 
 import logging
 import os
+from contextlib import contextmanager
 from pathlib import Path
 from typing import Optional, Union
 
 logger = logging.getLogger(__name__)
 
 
+@contextmanager
+def force_offline_if_cached(is_cached: bool, model_label: str = ""):
+    """Context manager that sets ``HF_HUB_OFFLINE=1`` while loading a cached model.
+
+    If *is_cached* is ``False`` the block runs normally (network allowed).
+    If the offline load raises an error containing "offline" we automatically
+    retry with network access so a partially-cached model still works.
+
+    Args:
+        is_cached: Whether the model weights are already on disk.
+        model_label: Human-readable name used in log messages.
+    """
+    if not is_cached:
+        yield
+        return
+
+    original_value = os.environ.get("HF_HUB_OFFLINE")
+    os.environ["HF_HUB_OFFLINE"] = "1"
+    logger.info(
+        "[offline-guard] %s is cached — forcing HF_HUB_OFFLINE=1",
+        model_label or "model",
+    )
+
+    try:
+        yield
+    except Exception as exc:
+        if "offline" in str(exc).lower():
+            logger.warning(
+                "[offline-guard] Offline load failed for %s, retrying with network: %s",
+                model_label or "model",
+                exc,
+            )
+            # Restore original env and retry — caller must wrap the load
+            # inside force_offline_if_cached so retrying here isn't possible.
+            # Instead, propagate a flag via the exception so the caller can
+            # decide.  For simplicity we just let it fall through to the
+            # finally block and re-raise.
+            raise
+        raise
+    finally:
+        if original_value is not None:
+            os.environ["HF_HUB_OFFLINE"] = original_value
+        else:
+            os.environ.pop("HF_HUB_OFFLINE", None)
+
+
 def patch_huggingface_hub_offline():
     """Monkey-patch huggingface_hub to force offline mode."""
     try:

backend/utils/progress.py

@@ -246,6 +246,8 @@ class ProgressManager:
                     # Send heartbeat
                     yield ": heartbeat\n\n"
                     continue
+        except (BrokenPipeError, ConnectionResetError, asyncio.CancelledError):
+            logger.debug(f"SSE client disconnected from {model_name}")
         finally:
             # Remove from listeners
             if model_name in self._listeners:

docs/content/docs/developer/building.mdx

@@ -159,12 +159,14 @@ Tauri looks for `voicebox-server-${PLATFORM}` in `src-tauri/binaries/` and bundl
 
 The `build-cuda-windows` job runs separately:
 
-1. Install PyTorch with CUDA 12.1
-2. Build with `build_binary.py --cuda`
-3. Split binary with `scripts/split_binary.py`
-4. Upload parts as release artifacts
+1. Install PyTorch with CUDA 12.8
+2. Build with `build_binary.py --cuda` (produces `--onedir` output)
+3. Package with `scripts/package_cuda.py` into two archives:
+   - `voicebox-server-cuda.tar.gz` — server core (~945 MB)
+   - `cuda-libs-cu128-v1.tar.gz` — NVIDIA runtime libraries (~1.7 GB, cached independently)
+4. Upload archives as release artifacts
 
-This binary is downloaded on-demand by users who enable CUDA in settings.
+This binary is downloaded on-demand by users who enable CUDA in settings. The CUDA libs archive is only re-downloaded when the CUDA toolkit version changes, not on every app update.
 
 ## Troubleshooting
 

docs/content/docs/developer/tts-engines.mdx

@@ -3,8 +3,12 @@ title: "TTS Engines"
 description: "How to add new text-to-speech engines to Voicebox"
 ---
 
+> **For humans:** This doc is optimized for AI agents to implement new TTS engines autonomously. It's structured as a phased workflow with explicit gates and a checklist so an agent can do the full integration — dependency research, backend, frontend, bundling — and hand you a draft release or prod build to test locally. It's also a useful reference if you're doing it yourself.
+
 Adding an engine touches ~10 files across 4 layers. The backend protocol work is straightforward — the real time sink is dependency hell, upstream library bugs, and PyInstaller bundling.
 
+**Do not start writing code until you complete Phase 0.** The v0.2.3 release was three patch releases of PyInstaller fixes because dependency research was skipped. Every issue — `inspect.getsource()` failures, missing native data files, metadata lookups, dtype mismatches — was discoverable by reading the model library's source code before integration began.
+
 ## Architecture Overview
 
 The backend is split into layers:
@@ -18,6 +22,133 @@ The backend is split into layers:
 
 New engines only need to touch `backends/` and `models.py` on the backend side — the route and service layers use a model config registry that handles dispatch automatically.
 
+## Phase 0: Dependency Research
+
+**This phase is mandatory.** Clone the model library and its key dependencies into a temporary directory and inspect them before writing any integration code. The goal is to produce a dependency audit that identifies every PyInstaller-incompatible pattern, every native data file, and every upstream bug you'll need to work around.
+
+### 0.1 Clone and Inspect the Model Library
+
+```bash
+# Create a throwaway workspace
+mkdir /tmp/engine-research && cd /tmp/engine-research
+
+# Clone the model library
+git clone https://github.com/org/model-library.git
+cd model-library
+```
+
+**Read these files first, in order:**
+
+1. **`setup.py` / `setup.cfg` / `pyproject.toml`** — Check pinned dependency versions. If the library pins `torch==2.6.0` or `numpy<1.26`, you'll need `--no-deps` installation and manual sub-dependency listing (this is what happened with `chatterbox-tts`).
+
+2. **`__init__.py` and the main model class** — Trace the import chain. Look for:
+   - `from_pretrained()` — does it call `huggingface_hub` internally? Does it pass `token=True` (which crashes without a stored HF token)?
+   - `from_local()` — does it exist? You may need manual `snapshot_download()` + `from_local()` to bypass download bugs.
+   - Device handling — does it default to CUDA? Does it support MPS? Many libraries crash on MPS with unsupported operators.
+
+3. **All `import` statements** — Recursively trace what the library imports. You're looking for:
+   - `inspect.getsource()` anywhere in the chain (search all `.py` files)
+   - `typeguard` / `@typechecked` decorators (these call `inspect.getsource()` at import time)
+   - `importlib.metadata.version()` or `pkg_resources.get_distribution()` (need `--copy-metadata`)
+   - `lazy_loader` (needs `--collect-all` to bundle `.pyi` stubs)
+
+### 0.2 Scan for PyInstaller-Incompatible Patterns
+
+Run these searches against the cloned library **and** its transitive dependencies:
+
+```bash
+# inspect.getsource — will crash in frozen binary without --collect-all
+grep -r "inspect.getsource\|getsource(" .
+
+# typeguard / @typechecked — calls inspect.getsource at import time
+grep -r "@typechecked\|from typeguard" .
+
+# importlib.metadata — needs --copy-metadata
+grep -r "importlib.metadata\|pkg_resources.get_distribution\|pkg_resources.require" .
+
+# Data files loaded at runtime — need --collect-all or --collect-data
+grep -r "Path(__file__).parent\|os.path.dirname(__file__)\|resources_path\|pkg_resources.resource_filename" .
+
+# Native library paths — may need env var override in frozen builds
+grep -r "/usr/share\|/usr/lib\|/usr/local\|espeak\|phonemize" .
+
+# torch.load without map_location — will crash on CPU-only builds
+grep -r "torch.load(" . | grep -v "map_location"
+
+# HuggingFace token bugs
+grep -r 'token=True\|token=os.getenv' .
+
+# Float64/Float32 assumptions — librosa returns float64, many models assume float32
+grep -r "torch.from_numpy\|\.double()\|float64" .
+
+# @torch.jit.script — calls inspect.getsource(), crashes in frozen builds
+grep -r "@torch.jit.script\|torch.jit.script" .
+
+# torchaudio.load — requires torchcodec in torchaudio 2.10+, use soundfile.read() instead
+grep -r "torchaudio.load\|torchaudio.save" .
+
+# Gated HuggingFace repos — models that hardcode gated repos as tokenizer/config sources
+grep -r "from_pretrained\|tokenizer_name\|AutoTokenizer" . | grep -i "llama\|meta-llama\|gated"
+```
+
+### 0.3 Install and Trace in a Throwaway Venv
+
+```bash
+# Create isolated venv
+python -m venv /tmp/engine-venv
+source /tmp/engine-venv/bin/activate
+
+# Install the package (try normally first)
+pip install model-package
+
+# Check if it conflicts with our stack
+pip install model-package torch==2.10 transformers==4.57.3 numpy>=1.26
+# If this fails, you need --no-deps:
+pip install --no-deps model-package
+
+# Get the full dependency tree
+pip show model-package  # Check Requires: field
+pip show -f model-package  # List all installed files (look for data files)
+
+# Check for non-PyPI dependencies
+pip install model-package 2>&1 | grep -i "no matching distribution"
+```
+
+### 0.4 Test Model Loading on CPU
+
+Before writing any integration code, verify the model works on CPU in a plain Python script:
+
+```python
+import torch
+# Force CPU to catch map_location bugs early
+model = ModelClass.from_pretrained("org/model", device="cpu")
+
+# Test with a float32 audio array (not float64)
+import numpy as np
+audio = np.random.randn(16000).astype(np.float32)
+output = model.generate("Hello world", audio)
+print(f"Output shape: {output.shape}, dtype: {output.dtype}, sample rate: {model.sample_rate}")
+```
+
+If this crashes, you've found a bug you'll need to monkey-patch. Common ones:
+- `RuntimeError: expected scalar type Float but found Double` → needs float32 cast
+- `RuntimeError: map_location` → needs `torch.load` patch
+- `RuntimeError: Unsupported operator aten::...` → needs MPS skip
+
+### 0.5 Produce a Dependency Audit
+
+Before proceeding to Phase 1, write down:
+
+1. **PyPI vs non-PyPI deps** — which packages need `--find-links`, `git+https://`, or `--no-deps`?
+2. **PyInstaller directives needed** — which packages need `--collect-all`, `--copy-metadata`, `--hidden-import`?
+3. **Runtime data files** — which packages ship data files (YAML, pretrained weights, phoneme tables, shader libraries) that must be bundled?
+4. **Native library paths** — which packages look for data at system paths that won't exist in a frozen binary?
+5. **Monkey-patches needed** — `torch.load` map_location, float64→float32 casts, MPS skip, HF token bypass, etc.
+6. **Sample rate** — what does the engine output? (24kHz, 44.1kHz, 48kHz)
+7. **Model download method** — `from_pretrained()` with library-managed download, or manual `snapshot_download()` + `from_local()`?
+
+This audit becomes your implementation plan for Phases 1, 4, and 5.
+
 ## Phase 1: Backend Implementation
 
 ### 1.1 Create the Backend File
@@ -148,6 +279,8 @@ In `app/src/lib/hooks/useGenerationForm.ts`:
 - Add engine-to-model-name mapping
 - Update payload construction for engine-specific fields
 
+**Watch out for model naming inconsistencies.** The HuggingFace repo name, the model size label, and the API model name don't always follow predictable patterns. For example, TADA's 3B model is named `tada-3b-ml` (not `tada-3b`), because it's a multilingual variant. Always check the actual repo names and build the frontend model name mapping from those, not from assumptions like `{engine}-{size}`.
+
 ### 3.5 Model Management
 
 In `app/src/components/ServerSettings/ModelManagement.tsx`:
@@ -155,54 +288,173 @@ In `app/src/components/ServerSettings/ModelManagement.tsx`:
 
 ## Phase 4: Dependencies
 
+Use the dependency audit from Phase 0 to drive this phase. You should already know what packages are needed, which conflict, and which require special installation.
+
 ### 4.1 Python Dependencies
 
-Add to `backend/requirements.txt`. Watch for:
+Add to `backend/requirements.txt`. There are three installation patterns, depending on what Phase 0 revealed:
 
-**Pinned dependency conflicts** — If the model package pins old versions, install with `--no-deps`:
+**Normal PyPI packages:**
+```
+some-model-package>=1.0.0
+```
+
+**Pinned dependency conflicts (`--no-deps`)** — If the model package pins old versions of torch/numpy/transformers, install with `--no-deps` and list sub-dependencies manually. This is the pattern used for `chatterbox-tts`:
 ```bash
+# In justfile / CI setup:
 pip install --no-deps chatterbox-tts
+
+# In requirements.txt — list each actual sub-dependency:
+conformer>=0.3.2
+diffusers>=0.31.0
+omegaconf>=2.3.0
+resemble-perth>=0.0.2
+s3tokenizer>=0.1.6
 ```
 
-Then list sub-dependencies manually in `requirements.txt`.
+To identify sub-deps: `pip show chatterbox-tts` → `Requires:` field, then cross-reference against existing `requirements.txt` to avoid duplicates.
 
-**Non-PyPI packages:**
-```
-linacodec @ git+https://github.com/user/repo.git
+**Non-PyPI packages** — Some libraries only exist on GitHub or require custom indexes:
 ```
+# Git-only packages (no PyPI release)
+linacodec @ git+https://github.com/ysharma3501/LinaCodec.git
+Zipvoice @ git+https://github.com/ysharma3501/LuxTTS.git
 
-**Custom package indexes:**
-```
+# Custom package indexes (C extensions with platform-specific wheels)
 --find-links https://k2-fsa.github.io/icefall/piper_phonemize.html
+piper-phonemize>=1.2.0
 ```
 
-### 4.2 Identifying Hidden Sub-Dependencies
+### 4.2 Dependency Conflict Resolution
 
-1. Install the package normally in a throwaway venv
-2. Run `pip show <package>` to get its `Requires:` list
-3. Cross-reference against existing requirements.txt
-4. Test that the engine loads and generates
+Check for conflicts with the existing stack before adding anything:
 
-## Phase 5: PyInstaller Bundling
+```bash
+# Our current stack pins (approximate):
+# Python 3.12+, torch>=2.10, transformers>=4.57, numpy>=1.26
 
-This is where most of the pain lives. Common issues:
+# Test compatibility
+pip install model-package torch==2.10 transformers==4.57.3 numpy>=1.26
 
-| Issue | Symptom | Fix |
-|-------|---------|-----|
-| `inspect.getsource()` at import | "could not get source code" | `--collect-all <package>` |
-| Data files (yaml, .pth.tar) | FileNotFoundError at runtime | `--collect-all <package>` |
-| Native data paths (espeak-ng) | Library looks at `/usr/share/...` | Set env var in frozen builds |
-| `importlib.metadata` lookups | "No package metadata found" | `--copy-metadata <package>` |
-| Dynamic imports | ModuleNotFoundError | `--hidden-import <module>` |
+# If it fails, check what the package pins:
+pip show model-package | grep Requires
+# Look at setup.py/pyproject.toml for version constraints
+```
 
-### Testing Frozen Builds
+**Known incompatible patterns in the wild:**
+- `torch==2.6.0` — many older packages pin this
+- `numpy<1.26` — conflicts with Python 3.12+
+- `transformers==4.46.3` — many packages pin old transformers
+- `onnxruntime` pinned versions — often conflict with torch
 
-You can't skip this. Models that work in `python -m uvicorn` will break in the PyInstaller binary.
+### 4.3 Update Installation Scripts
+
+Dependencies must be added in multiple places:
+
+| File | What to add |
+|------|------------|
+| `backend/requirements.txt` | Package and version constraint |
+| `justfile` | `--no-deps` install line if needed (in `setup-python` and `setup-python-release` targets) |
+| `.github/workflows/release.yml` | Same `--no-deps` line in CI build steps |
+| `Dockerfile` | Same install commands for Docker builds |
+
+## Phase 5: PyInstaller Bundling (`build_binary.py`)
+
+This is where most of the pain lives. **The v0.2.3 release was entirely dedicated to fixing bundling issues** — every new engine that shipped in v0.2.1 (LuxTTS, Chatterbox, Chatterbox Turbo) worked in dev but failed in production builds. Don't skip this phase.
+
+### 5.1 Register Your Engine in `build_binary.py`
+
+Every new engine needs entries in `backend/build_binary.py`. This file drives PyInstaller and is the single most common source of "works in dev, breaks in prod" bugs. You need to decide which PyInstaller directives your engine's dependencies require:
+
+| Directive | What It Does | When You Need It |
+|-----------|-------------|-----------------|
+| `--hidden-import <module>` | Includes a module PyInstaller can't detect via static analysis | Dynamic imports, lazy imports, plugin architectures |
+| `--collect-all <package>` | Bundles source `.py` files, data files, AND native libraries | Packages that call `inspect.getsource()` at import time (e.g. `inflect` via `typeguard`'s `@typechecked`), or that ship pretrained model files (e.g. `perth` ships `.pth.tar` + `hparams.yaml`) |
+| `--collect-data <package>` | Bundles only data files (not source or native libs) | Packages with YAML configs, vocab files, etc. |
+| `--collect-submodules <package>` | Bundles all submodules | Packages with deep module trees that PyInstaller misses |
+| `--copy-metadata <package>` | Copies `importlib.metadata` info | Packages that call `importlib.metadata.version()` or `pkg_resources.get_distribution()` at runtime. Already required for: `requests`, `transformers`, `huggingface-hub`, `tokenizers`, `safetensors`, `tqdm` |
+
+**Example: adding hidden imports and collect-all for a new engine:**
+
+```python
+# In build_binary.py, inside the args list:
+"--hidden-import",
+"backend.backends.your_engine_backend",
+"--hidden-import",
+"your_engine_package",
+"--hidden-import",
+"your_engine_package.inference",
+"--collect-all",
+"some_dependency_that_uses_inspect_getsource",
+"--copy-metadata",
+"some_dependency_that_checks_its_own_version",
+```
+
+### 5.2 Lessons from v0.2.3 — Real Failures and Their Fixes
+
+These are actual production failures from shipping new engines. Every one of these passed `python -m uvicorn` in dev:
+
+| Engine | Failure | Root Cause | Fix |
+|--------|---------|-----------|-----|
+| LuxTTS | `"could not get source code"` on import | `inflect` uses `typeguard`'s `@typechecked` which calls `inspect.getsource()` — needs `.py` source files, not just bytecode | `--collect-all inflect` |
+| LuxTTS | `espeak-ng-data` not found | `piper_phonemize` C library looks for data at `/usr/share/espeak-ng-data/` which doesn't exist in the bundle | `--collect-all piper_phonemize` + set `ESPEAK_DATA_PATH` env var at runtime (see 5.3) |
+| LuxTTS | `inspect.getsource` error in Vocos codec | `linacodec` and `zipvoice` use source introspection | `--collect-all linacodec` + `--collect-all zipvoice` |
+| Chatterbox | `FileNotFoundError` for watermark model | `perth` ships pretrained model files (`hparams.yaml`, `.pth.tar`) that PyInstaller doesn't bundle by default | `--collect-all perth` |
+| All engines | `importlib.metadata` failures | Frozen binary doesn't include package metadata for `huggingface-hub`, `transformers`, etc. | `--copy-metadata` for each affected package |
+| All engines | Download progress bars stuck at 0% | `huggingface_hub` silently disables tqdm progress bars based on logger level in frozen builds — our progress tracker never receives byte updates | Force-enable tqdm's internal counter in `HFProgressTracker` |
+| TADA | `inspect.getsource` error in DAC's `Snake1d` | `@torch.jit.script` calls `inspect.getsource()` which fails without `.py` source files | Wrote a lightweight shim (`dac_shim.py`) reimplementing `Snake1d` without `@torch.jit.script`, registered fake `dac.*` modules in `sys.modules` |
+| All engines | `NameError: name 'obj' is not defined` on macOS | Python 3.12.0 has a [CPython bug](https://github.com/pyinstaller/pyinstaller/issues/7992) that corrupts bytecode when PyInstaller rewrites code objects | Upgrade to Python 3.12.13+ |
+| All engines | `resource_tracker` subprocess crash | `multiprocessing` in frozen binaries needs `freeze_support()` called before anything else | Added to `server.py` entry point |
+
+### 5.3 Runtime Frozen-Build Handling (`server.py`)
+
+Some fixes can't live in `build_binary.py` — they need runtime detection. The entry point `backend/server.py` handles these before any heavy imports:
+
+```python
+# 1. freeze_support() — MUST be called before any multiprocessing use
+import multiprocessing
+multiprocessing.freeze_support()
+
+# 2. Native data paths — redirect C libraries to bundled data
+if getattr(sys, 'frozen', False):
+    _meipass = getattr(sys, '_MEIPASS', os.path.dirname(sys.executable))
+    _espeak_data = os.path.join(_meipass, 'piper_phonemize', 'espeak-ng-data')
+    if os.path.isdir(_espeak_data):
+        os.environ.setdefault('ESPEAK_DATA_PATH', _espeak_data)
+
+# 3. stdout/stderr safety — PyInstaller --noconsole on Windows sets these to None
+if not _is_writable(sys.stdout):
+    sys.stdout = open(os.devnull, 'w')
+```
+
+If your engine's dependencies include native libraries that look for data at system paths (like espeak-ng does), you'll need to add a similar `os.environ.setdefault()` block here.
+
+### 5.4 CUDA vs CPU Build Branching
+
+`build_binary.py` produces two different binaries:
+
+- **`voicebox-server`** (CPU) — excludes all `nvidia.*` packages to avoid bundling ~3 GB of CUDA DLLs
+- **`voicebox-server-cuda`** — includes `torch.cuda` and `torch.backends.cudnn`
+
+On Windows, if the build environment has CUDA torch installed but you're building the CPU binary, the script temporarily swaps to CPU-only torch and restores CUDA torch afterward. This prevents PyInstaller from accidentally bundling CUDA libraries into the CPU build.
+
+New engine imports go in the **common section** (not the CUDA or MLX conditional blocks) unless your engine has platform-specific dependencies.
+
+### 5.5 MLX Conditional Inclusion
+
+Apple Silicon builds conditionally include MLX hidden imports and `--collect-all mlx` / `--collect-all mlx_audio`. If your engine has an MLX-specific backend variant, add its imports inside the `if is_apple_silicon() and not cuda:` block.
+
+### 5.6 Testing Frozen Builds
+
+You can't skip this. Models that work in `python -m uvicorn` will break in the PyInstaller binary. The v0.2.3 release required **three patch releases** (v0.2.1 → v0.2.2 → v0.2.3) to get all engines working in production.
 
 1. Build: `just build`
-2. Run and try download + load + generate
-3. Check stderr for the actual error
-4. Fix, rebuild, repeat
+2. Launch the binary directly (not via `python -m`)
+3. Test the **full chain**: download → load → generate → progress tracking
+4. Check stderr for the actual error (logs go to stderr for Tauri sidecar capture)
+5. Fix, rebuild, repeat
+
+**Common gotcha:** testing only generation with a pre-cached model from your dev install. Always test with a clean model cache to verify downloads work too.
 
 ## Phase 6: Common Upstream Workarounds
 
@@ -240,6 +492,90 @@ def _get_device(self):
     return "cpu"  # Skip MPS
 ```
 
+### Gated HuggingFace repos as hardcoded config sources
+
+Some models hardcode a gated HuggingFace repo as their tokenizer or config source (e.g., TADA hardcodes `"meta-llama/Llama-3.2-1B"` in both its `AlignerConfig` and `TadaConfig`). This silently fails without HF authentication.
+
+**Fix:** Download from an ungated mirror and patch the config objects directly:
+
+```python
+# Download tokenizer from ungated mirror
+UNGATED_TOKENIZER = "unsloth/Llama-3.2-1B"
+tokenizer_path = snapshot_download(UNGATED_TOKENIZER, token=None)
+
+# Patch the model config to use the local path instead of the gated repo
+config = ModelConfig.from_pretrained(model_path)
+config.tokenizer_name = tokenizer_path
+model = ModelClass.from_pretrained(model_path, config=config)
+```
+
+**Do NOT monkey-patch `AutoTokenizer.from_pretrained`** — it's a classmethod, and replacing it corrupts the descriptor, which breaks other engines that use different tokenizers (e.g., Qwen uses a Qwen tokenizer via `AutoTokenizer`). Always patch at the config level, not the class method level.
+
+### `torchaudio.load()` requires `torchcodec` in 2.10+
+
+As of `torchaudio>=2.10`, `torchaudio.load()` requires the `torchcodec` package for audio I/O. If your engine or backend code uses `torchaudio.load()`, replace it with `soundfile`:
+
+```python
+# Before (breaks without torchcodec):
+import torchaudio
+waveform, sr = torchaudio.load("audio.wav")
+
+# After:
+import soundfile as sf
+import torch
+data, sr = sf.read("audio.wav", dtype="float32")
+waveform = torch.from_numpy(data).unsqueeze(0)
+```
+
+Note: `torchaudio.functional.resample()` and other pure-PyTorch math functions work fine without `torchcodec` — only the I/O functions are affected.
+
+### `@torch.jit.script` breaks in frozen builds
+
+`torch.jit.script` calls `inspect.getsource()` to parse the decorated function's source code. In a PyInstaller binary, `.py` source files aren't available, so this crashes at import time.
+
+**Fix:** Remove or avoid `@torch.jit.script` decorators. If the decorated function comes from an upstream dependency, write a shim that reimplements the function without the decorator (see "Toxic dependency chains" below).
+
+### Toxic dependency chains — the shim pattern
+
+Sometimes a model library depends on a package with a massive, hostile transitive dependency tree, but only uses a tiny piece of it. When the dependency chain is unbuildable or would pull in dozens of unwanted packages, the right move is to write a lightweight shim.
+
+**Example:** TADA depends on `descript-audio-codec` (DAC), which pulls in `descript-audiotools` -> `onnx`, `tensorboard`, `protobuf`, `matplotlib`, `pystoi`, etc. The `onnx` package fails to build from source on macOS. But TADA only uses `Snake1d` from DAC — a 7-line PyTorch module.
+
+**Solution:** Create a shim at `backend/utils/dac_shim.py` that registers fake modules in `sys.modules`:
+
+```python
+import sys
+import types
+import torch
+from torch import nn
+
+def snake(x, alpha):
+    """Snake activation — reimplemented without @torch.jit.script."""
+    return x + (1.0 / (alpha + 1e-9)) * torch.sin(alpha * x).pow(2)
+
+class Snake1d(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.alpha = nn.Parameter(torch.ones(1, channels, 1))
+    def forward(self, x):
+        return snake(x, self.alpha)
+
+# Register fake dac.* modules so "from dac.nn.layers import Snake1d" works
+_nn = types.ModuleType("dac.nn")
+_layers = types.ModuleType("dac.nn.layers")
+_layers.Snake1d = Snake1d
+_nn.layers = _layers
+
+for name, mod in [("dac", types.ModuleType("dac")),
+                   ("dac.nn", _nn), ("dac.nn.layers", _layers)]:
+    sys.modules[name] = mod
+```
+
+**Key rules for shims:**
+- Import the shim **before** importing the model library (so it finds the fake modules first)
+- Do NOT use `@torch.jit.script` in the shim (see above)
+- Only reimplement what the model actually uses — check the import chain carefully
+
 ## Upcoming Engines
 
 Based on the current model landscape, these are candidates for future integration:
@@ -250,8 +586,83 @@ Based on the current model landscape, these are candidates for future integratio
 | **Fish Speech** | 50+ | Medium | Word-level control via inline text | Ready |
 | **Kokoro-82M** | English | 82M | CPU realtime, Apache 2.0 | Ready |
 | **XTTS-v2** | 17+ | Medium | Zero-shot cloning | Ready |
-| **HumeAI TADA** | EN (1B), Multi (3B) | Medium | 700s+ coherent audio, synced transcripts | Needs vetting |
 | **MOSS-TTS** | Multilingual | Medium | Text-to-voice design, multi-speaker dialogue | Needs vetting |
 | **Pocket TTS** | English | ~100M | CPU-first, >1× realtime | Needs vetting |
 
-The multi-engine architecture is now in place, making new model integration straightforward (~1 day for a well-documented model with a PyPI package).
+## Implementation Checklist
+
+Use this as a gate between phases. Do not proceed to the next phase until every item in the current phase is checked.
+
+### Phase 0: Dependency Research
+- [ ] Cloned model library source into a temp directory
+- [ ] Read `setup.py` / `pyproject.toml` — noted pinned dependency versions
+- [ ] Traced all imports from the model class through to leaf dependencies
+- [ ] Searched for `inspect.getsource`, `@typechecked`, `typeguard` in the full dependency tree
+- [ ] Searched for `importlib.metadata`, `pkg_resources.get_distribution` in the dependency tree
+- [ ] Searched for `Path(__file__).parent`, `os.path.dirname(__file__)`, hardcoded system paths
+- [ ] Searched for `torch.load` calls missing `map_location`
+- [ ] Searched for `torch.from_numpy` without `.float()` cast
+- [ ] Searched for `token=True` or `token=os.getenv("HF_TOKEN")` in HuggingFace calls
+- [ ] Searched for `@torch.jit.script` / `torch.jit.script` (crashes in frozen builds)
+- [ ] Searched for `torchaudio.load` / `torchaudio.save` (requires `torchcodec` in 2.10+)
+- [ ] Searched for hardcoded gated HuggingFace repo names (e.g., `meta-llama/*`)
+- [ ] Evaluated whether any dependency is used minimally enough to shim instead of install
+- [ ] Tested model loading and generation on CPU in a throwaway venv
+- [ ] Tested with a clean HuggingFace cache (no pre-downloaded models)
+- [ ] Produced a written dependency audit documenting all findings
+
+### Phase 1: Backend Implementation
+- [ ] Created `backend/backends/<engine>_backend.py` implementing `TTSBackend` protocol
+- [ ] Chose voice prompt pattern (pre-computed tensors vs deferred file paths)
+- [ ] Implemented all monkey-patches identified in Phase 0
+- [ ] Used `get_torch_device()` from `backends/base.py` for device selection
+- [ ] Used `model_load_progress()` from `backends/base.py` for download/load tracking
+- [ ] Tested: model downloads correctly
+- [ ] Tested: model loads on CPU
+- [ ] Tested: generation produces valid audio
+- [ ] Tested: voice cloning from reference audio works
+- [ ] Registered `ModelConfig` in `backends/__init__.py`
+- [ ] Added to `TTS_ENGINES` dict
+- [ ] Added factory branch in `get_tts_backend_for_engine()`
+- [ ] Updated engine regex in `backend/models.py`
+
+### Phase 2–3: Route, Service, and Frontend
+- [ ] Confirmed zero changes needed in routes/services (or documented why custom behavior is needed)
+- [ ] Added engine to TypeScript union type in `app/src/lib/api/types.ts`
+- [ ] Added language map entry in `app/src/lib/constants/languages.ts`
+- [ ] Added to `ENGINE_OPTIONS` and `ENGINE_DESCRIPTIONS` in `EngineModelSelector.tsx`
+- [ ] Added to Zod schema and model-name mapping in `useGenerationForm.ts`
+- [ ] Added description in `ModelManagement.tsx`
+
+### Phase 4: Dependencies
+- [ ] Added packages to `backend/requirements.txt`
+- [ ] If `--no-deps` needed: listed sub-dependencies explicitly
+- [ ] If git-only packages: added `@ git+https://...` entries
+- [ ] If custom index needed: added `--find-links` line
+- [ ] Updated `justfile` setup targets
+- [ ] Updated `.github/workflows/release.yml` build steps
+- [ ] Updated `Dockerfile` if applicable
+- [ ] Verified `pip install` succeeds in a clean venv with existing requirements
+
+### Phase 5: PyInstaller Bundling
+- [ ] Added `--hidden-import` entries in `build_binary.py` for:
+  - [ ] `backend.backends.<engine>_backend`
+  - [ ] The model package and its key submodules
+- [ ] Added `--collect-all` for any packages that:
+  - [ ] Use `inspect.getsource()` / `@typechecked`
+  - [ ] Ship pretrained model data files (`.pth.tar`, `.yaml`, etc.)
+  - [ ] Ship native data files (phoneme tables, shader libraries, etc.)
+- [ ] Added `--copy-metadata` for any packages that use `importlib.metadata`
+- [ ] If engine has native data paths: added `os.environ.setdefault()` in `server.py`
+- [ ] Built frozen binary with `just build`
+- [ ] Tested in frozen binary with **clean model cache** (not pre-cached from dev):
+  - [ ] Model download works with real-time progress
+  - [ ] Model loading works
+  - [ ] Generation produces valid audio
+  - [ ] No errors in stderr logs
+
+### Phase 6: Final Verification
+- [ ] Engine works in dev mode (`just dev`)
+- [ ] Engine works in frozen binary (`just build` → run binary directly)
+- [ ] Tested on target platform (macOS for MLX, Windows/Linux for CUDA)
+- [ ] No regressions in existing engines

docs/content/docs/index.mdx

@@ -3,12 +3,12 @@ title: "Voicebox Documentation"
 description: "Voicebox is a local-first voice cloning studio -- a free and open-source alternative to ElevenLabs."
 ---
 
-Voicebox is a **local-first voice cloning studio** -- a free and open-source alternative to ElevenLabs. Clone voices from a few seconds of audio, generate speech in 23 languages across 4 TTS engines, apply post-processing effects, and compose multi-voice projects with a timeline editor.
+Voicebox is a **local-first voice cloning studio** -- a free and open-source alternative to ElevenLabs. Clone voices from a few seconds of audio, generate speech in 23 languages across 5 TTS engines, apply post-processing effects, and compose multi-voice projects with a timeline editor.
 
 ![Voicebox App Screenshot](/images/app-screenshot-1.webp)
 
 - **Complete privacy** -- models and voice data stay on your machine
-- **4 TTS engines** -- Qwen3-TTS, LuxTTS, Chatterbox Multilingual, and Chatterbox Turbo
+- **5 TTS engines** -- Qwen3-TTS, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, and HumeAI TADA
 - **23 languages** -- from English to Arabic, Japanese, Hindi, Swahili, and more
 - **Post-processing effects** -- pitch shift, reverb, delay, chorus, compression, and filters
 - **Expressive speech** -- paralinguistic tags like `[laugh]`, `[sigh]`, `[gasp]` via Chatterbox Turbo

docs/content/docs/overview/introduction.mdx

@@ -5,10 +5,10 @@ description: "Voicebox is a local-first voice cloning studio -- a free and open-
 
 ## What is Voicebox?
 
-Voicebox is a **local-first voice cloning studio** -- a free and open-source alternative to ElevenLabs. Clone voices from a few seconds of audio, generate speech in 23 languages across 4 TTS engines, apply post-processing effects, and compose multi-voice projects with a timeline editor.
+Voicebox is a **local-first voice cloning studio** -- a free and open-source alternative to ElevenLabs. Clone voices from a few seconds of audio, generate speech in 23 languages across 5 TTS engines, apply post-processing effects, and compose multi-voice projects with a timeline editor.
 
 - **Complete privacy** -- models and voice data stay on your machine
-- **4 TTS engines** -- Qwen3-TTS, LuxTTS, Chatterbox Multilingual, and Chatterbox Turbo
+- **5 TTS engines** -- Qwen3-TTS, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, and HumeAI TADA
 - **23 languages** -- from English to Arabic, Japanese, Hindi, Swahili, and more
 - **Post-processing effects** -- pitch shift, reverb, delay, chorus, compression, and filters
 - **Expressive speech** -- paralinguistic tags like `[laugh]`, `[sigh]`, `[gasp]` via Chatterbox Turbo
@@ -20,7 +20,7 @@ Voicebox is a **local-first voice cloning studio** -- a free and open-source alt
 
 ## TTS Engines
 
-Four engines with different strengths, switchable per-generation:
+Five engines with different strengths, switchable per-generation:
 
 | Engine | Languages | Strengths |
 |--------|-----------|-----------|
@@ -28,6 +28,7 @@ Four engines with different strengths, switchable per-generation:
 | **LuxTTS** | English | Lightweight (~1GB VRAM), 48kHz output, 150x realtime on CPU |
 | **Chatterbox Multilingual** | 23 | Broadest language coverage |
 | **Chatterbox Turbo** | English | Fast 350M model with paralinguistic emotion/sound tags |
+| **TADA** (1B / 3B) | 10 | HumeAI speech-language model -- 700s+ coherent audio |
 
 ## GPU Support
 
@@ -56,7 +57,7 @@ Four engines with different strengths, switchable per-generation:
 | Frontend | React, TypeScript, Tailwind CSS |
 | State | Zustand, React Query |
 | Backend | FastAPI (Python) |
-| TTS Engines | Qwen3-TTS, LuxTTS, Chatterbox, Chatterbox Turbo |
+| TTS Engines | Qwen3-TTS, LuxTTS, Chatterbox, Chatterbox Turbo, TADA |
 | Effects | Pedalboard (Spotify) |
 | Transcription | Whisper / Whisper Turbo (PyTorch or MLX) |
 | Inference | MLX (Apple Silicon) / PyTorch (CUDA/ROCm/XPU/CPU) |

docs/notes/PROJECT_STATUS.md

@@ -36,6 +36,10 @@
 │  │  │  │ Qwen3-TTS│ │LuxTTS │ │Chatterbox │  │  │  │
 │  │  │  │(Py/MLX)  │ │       │ │(MTL+Turbo)│  │  │  │
 │  │  │  └──────────┘ └───────┘ └───────────┘  │  │  │
+│  │  │  ┌──────────┐                           │  │  │
+│  │  │  │ TADA     │                           │  │  │
+│  │  │  │(1B / 3B) │                           │  │  │
+│  │  │  └──────────┘                           │  │  │
 │  │  └─────────────────────────────────────────┘  │  │
 │  │  ┌───────────┐  ┌─────────┐                   │  │
 │  │  │ STTBackend│  │ Profiles│                   │  │
@@ -59,6 +63,7 @@
 | 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 |
+| TADA | `backend/backends/hume_backend.py` | HumeAI TADA — 1B English + 3B Multilingual |
 | 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) |
@@ -78,7 +83,7 @@
 ```
 POST /generate
   1. Look up voice profile from DB
-  2. Resolve engine from request (qwen | luxtts | chatterbox | chatterbox_turbo)
+  2. Resolve engine from request (qwen | luxtts | chatterbox | chatterbox_turbo | tada)
   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)
@@ -104,7 +109,8 @@ POST /generate
 - LuxTTS integration — fast, CPU-friendly English TTS (PR #254)
 - Chatterbox Multilingual TTS — 23 languages including Hebrew (PR #257)
 - Instruct parameter UI exists but is non-functional across all backends (see #224, Known Limitations)
-- Single flat model dropdown (Qwen 1.7B, Qwen 0.6B, LuxTTS, Chatterbox, Chatterbox Turbo)
+- HumeAI TADA integration — 1B English + 3B Multilingual speech-language model (PR #296)
+- Single flat model dropdown (Qwen 1.7B, Qwen 0.6B, LuxTTS, Chatterbox, Chatterbox Turbo, TADA 1B, TADA 3B)
 - Centralized model config registry (`ModelConfig` dataclass) — no per-engine dispatch maps in `main.py`
 - Shared `EngineModelSelector` component — engine/model dropdown defined once, used in both generation forms
 
@@ -136,6 +142,8 @@ POST /generate
 | LuxTTS | `luxtts` | English | ~300 MB | CPU-friendly, 48 kHz, fast | None |
 | Chatterbox | `chatterbox-tts` | 23 (incl. Hebrew, Arabic, Hindi, etc.) | ~3.2 GB | Zero-shot cloning, multilingual | Partial — `exaggeration` float (0-1) for expressiveness |
 | Chatterbox Turbo | `chatterbox-turbo` | English | ~1.5 GB | Paralinguistic tags ([laugh], [cough]), 350M params, low latency | Partial — inline tags only, no separate instruct param |
+| TADA 1B | `tada-1b` | English | ~4 GB | HumeAI speech-language model, 700s+ coherent audio | None |
+| TADA 3B Multilingual | `tada-3b-ml` | 10 (en, ar, zh, de, es, fr, it, ja, pl, pt) | ~8 GB | Multilingual, text-acoustic dual alignment | None |
 
 ### Multi-Engine Architecture (Shipped)
 
@@ -143,7 +151,7 @@ The singleton TTS backend blocker described in the previous version of this doc
 
 - **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'`
+- **Engine field on GenerationRequest** — frontend sends `engine: 'qwen' | 'luxtts' | 'chatterbox' | 'chatterbox_turbo' | 'tada'`
 - **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)
@@ -337,7 +345,7 @@ Notable requests:
 | **CosyVoice2-0.5B** | 3-10s zero-shot | Very fast | 24 kHz | Multilingual | Low | **Yes** — `inference_instruct2()`, works with cloning | Ready | Best instruct candidate |
 | **Fish Speech** | 10-30s few-shot | Real-time | 24-44 kHz | 50+ | Medium | **Yes** — inline text descriptions, word-level control | Ready | Multi-engine arch in place |
 | **MOSS-TTS Family** | Zero-shot | — | — | Multilingual | Medium | **Yes** — text prompts for style + timbre design | Needs vetting | Apache 2.0, multi-speaker dialogue |
-| **HumeAI TADA 1B/3B** | Zero-shot | 5× faster than LLM-TTS | — | EN (1B), Multilingual (3B) | Medium | Partial — automatic prosody from text context | Needs vetting | MIT, 700s+ coherent, synced transcript output |
+| **HumeAI TADA 1B/3B** | Zero-shot | 5× faster than LLM-TTS | 24 kHz | EN (1B), Multilingual (3B) | Medium | Partial — automatic prosody from text context | **Shipped** | PR #296, MIT, 700s+ coherent |
 | **VoxCPM 1.5** | Zero-shot (seconds) | ~0.15 RTF streaming | — | Bilingual (EN/ZH) | Medium | Partial — automatic context-aware prosody | Needs vetting | Apache 2.0, tokenizer-free continuous diffusion |
 | **Kokoro-82M** | 3s instant | CPU realtime | 24 kHz | English | Tiny (82M) | Partial — automatic style inference | Ready | Apache 2.0, multi-engine arch in place |
 | **XTTS-v2** | 6s zero-shot | Mid-GPU | 24 kHz | 17+ | Medium | Partial — style transfer from ref audio only | Ready | Multi-engine arch in place |
@@ -475,7 +483,7 @@ The generation form now uses a flat model dropdown with engine-based routing. Pe
 | `/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 (Qwen, LuxTTS, Chatterbox, Chatterbox Turbo, Whisper) |
+| `/models/status` | GET | All model statuses (Qwen, LuxTTS, Chatterbox, Chatterbox Turbo, TADA, Whisper) |
 | `/models/download` | POST | Trigger model download |
 | `/models/download/cancel` | POST | Cancel/dismiss download |
 | `/models/{name}` | DELETE | Delete downloaded model |

docs/plans/CUDA_LIBS_ADDON.md

@@ -0,0 +1,173 @@
+# CUDA Libs as a Bolt-On Addon
+
+## Problem
+
+Every time we bump `__version__` (even for a UI tweak or bugfix), the exact-match version check in both `main.rs:222` and `cuda.py:237` invalidates the user's ~2.4GB CUDA binary, forcing a full redownload. The CUDA binary is the entire server rebuilt with NVIDIA libs included -- there's no separation between app logic and the CUDA runtime.
+
+## Why This Is Hard With `--onefile`
+
+The core tension is PyInstaller `--onefile` mode (`build_binary.py:39`). In onefile mode, everything -- Python code, all dependencies, torch, the NVIDIA `.dll`/`.so` files -- gets packed into a single self-extracting archive. There's no concept of "swap out one part." The binary IS the server.
+
+## Options
+
+### Option A: Switch to `--onedir` for the CUDA Build (Recommended)
+
+Instead of `--onefile`, build the CUDA variant as a directory (a folder with the exe + all the shared libs alongside it). Then split the distribution into two archives:
+
+1. **`voicebox-server-cuda` executable + non-NVIDIA deps** (~200-400MB) -- versioned with the app, redownloaded on every app update.
+2. **`cuda-libs-cu126.tar.gz`** (~2GB) -- the `nvidia.*` packages (cublas, cudnn, cuda_runtime, etc.), versioned independently (e.g., `cuda-libs-cu126-v1`). Only redownloaded when we bump the CUDA toolkit version or torch's CUDA dependency changes.
+
+#### How it would work at runtime
+
+- Tauri downloads the server binary archive and extracts it to `{data_dir}/backends/cuda/`
+- On first CUDA setup (or when cuda-libs version bumps), downloads and extracts the libs archive into the same directory
+- The CUDA server exe finds the `.dll`/`.so` files next to it (standard PyInstaller onedir behavior)
+- Version check becomes two checks: server version + cuda-libs version
+
+#### Independent versioning
+
+Add a `cuda-libs.json` manifest:
+
+```json
+{"version": "cu126-v1", "torch_compat": ">=2.6.0,<2.8.0"}
+```
+
+The server checks this on startup. The Tauri side checks it before launching. Only bump `cu126-v1` -> `cu126-v2` when we actually change the CUDA toolkit or torch major version.
+
+#### Build pipeline changes
+
+The CI `build-cuda-windows` job would build with `--onedir`, then separate the output into two archives. The CUDA libs archive could be built less frequently (only when torch/CUDA version changes) and stored as a pinned release asset.
+
+#### Download experience
+
+- First-time CUDA setup: ~2.4GB total (same as today)
+- Subsequent app updates: ~200-400MB for the server, CUDA libs stay cached
+- CUDA toolkit bump: ~2GB for just the libs
+
+#### Pros
+
+- PyInstaller `--onedir` natively produces this structure -- NVIDIA DLLs end up as discrete files in the output directory
+- The separation is natural: PyInstaller puts torch's NVIDIA deps in predictable paths (`nvidia/cublas/lib/`, etc.)
+- CUDA libs are highly stable -- only rebundle when changing CUDA toolkit version (e.g., cu126 -> cu128) or major torch version
+- Server updates become ~200-400MB instead of ~2.4GB
+- No library path hacking needed -- torch finds NVIDIA DLLs because they're in the same directory tree
+
+#### Cons
+
+- Onedir means a folder with hundreds of files instead of a single exe -- more complex to manage, extract, and clean up
+- Need to modify download/assembly logic in `cuda.py` to handle two separate archives
+- The Tauri side (`main.rs`) needs to point at an exe inside a directory rather than a standalone binary
+- Users who manually manage the file may find the folder structure confusing
+
+#### TTS engine compatibility
+
+No issues. The TTS engines are pure Python + torch. They don't care whether NVIDIA libs are inside the binary or sitting next to it -- torch's dynamic loader finds them either way.
+
+---
+
+### Option B: Keep `--onefile` but Externalize CUDA Libs via Library Path
+
+Keep the server as a single `--onefile` binary (with NVIDIA packages excluded, same as the CPU build). Ship the CUDA libs as a separate download that gets extracted to `{data_dir}/backends/cuda-libs/`. Before launching, set the library search path to include that directory.
+
+**Important caveat:** The CPU torch wheel (`whl/cpu`) doesn't have CUDA kernels compiled in -- it's a fundamentally different build. So the binary would need to be built with CUDA-compiled torch but with the NVIDIA runtime libraries excluded. The runtime libs (cublas, cudnn, etc.) would be provided externally.
+
+#### How it would work
+
+- Build ONE "CUDA-ready" server binary with CUDA-compiled torch but NVIDIA runtime packages excluded
+- Ship `cuda-libs-cu126-v1.tar.gz` separately (~2GB of `.dll`/`.so` files)
+- When launching, Tauri sets `PATH` (Windows) or `LD_LIBRARY_PATH` (Linux) to include the cuda-libs directory
+
+#### Pros
+
+- Single server binary for both CPU and CUDA users -- simplifies build pipeline enormously
+- True bolt-on CUDA libs with fully independent versioning
+- Server updates are always small (~150MB for the onefile binary)
+
+#### Cons
+
+- **Fragile on Windows.** PyInstaller `--onefile` extracts to a temp directory at runtime and the internal torch may not find externally-placed NVIDIA libs. DLL resolution on Windows is notoriously unreliable in this scenario.
+- `os.add_dll_directory()` only affects `LoadLibraryEx` with `LOAD_LIBRARY_SEARCH_USER_DIRS` flag -- not all DLL loads go through this path
+- PyInstaller's onefile bootloader may configure DLL search paths before Python code runs
+- Could work on Linux but is fragile on Windows
+
+---
+
+### Option C: Hybrid -- `--onefile` Server + Dynamic CUDA Lib Loading at Runtime
+
+Build the server as `--onefile` with CUDA-compiled torch but with NVIDIA packages excluded. At startup, before torch initializes CUDA, explicitly load the NVIDIA shared libraries using `ctypes.CDLL` or `os.add_dll_directory()`.
+
+In `server.py`, before any torch imports:
+
+```python
+cuda_libs_dir = os.environ.get("VOICEBOX_CUDA_LIBS")
+if cuda_libs_dir and os.path.isdir(cuda_libs_dir):
+    if sys.platform == "win32":
+        os.add_dll_directory(cuda_libs_dir)
+        os.environ["PATH"] = cuda_libs_dir + os.pathsep + os.environ.get("PATH", "")
+    else:
+        os.environ["LD_LIBRARY_PATH"] = cuda_libs_dir + ":" + os.environ.get("LD_LIBRARY_PATH", "")
+```
+
+#### Pros
+
+- Single server binary, true bolt-on CUDA libs
+- Clean separation of concerns
+- Independent versioning
+
+#### Cons
+
+- Needs careful testing with each torch version -- CUDA initialization happens deep in C++ extension layer
+- On Windows, `os.add_dll_directory()` may not cover all DLL load paths
+- PyInstaller's onefile bootloader may have already configured DLL search paths before Python code runs
+- Most complex to get right and maintain
+
+## Recommendation
+
+**Option A (`--onedir` with split archives)** is the most reliable path:
+
+1. **It actually works.** `--onedir` puts all files on disk as regular files. Torch finds NVIDIA DLLs because they're in the same directory tree, exactly as they would be in a normal pip install.
+2. **Natural separation.** PyInstaller's `--onedir` output already separates the NVIDIA `.dll`/`.so` files into `nvidia/` subdirectories. We can split the output directory into "core" and "nvidia-libs" archives after building.
+3. **Independent versioning is straightforward.** A `cuda-libs.json` manifest controls when redownloads are needed.
+4. **Build pipeline simplification.** Build CUDA libs archive less frequently, store as a pinned release asset.
+
+The main cost is managing a directory instead of a single file, but we already have sophisticated download/assembly infrastructure in `cuda.py` with manifests and split parts. Extending that to handle two archives is incremental work.
+
+## Tauri Compatibility (Validated)
+
+Tauri handles PyInstaller `--onedir` with no issues. The key insight is that we're **not** using a static sidecar for CUDA -- we're downloading and extracting at runtime (the existing `cuda.py` + `main.rs` flow). For runtime-launched processes, Tauri's `tauri::shell::Command` supports arbitrary directories natively.
+
+### The critical change in `main.rs`
+
+The only Tauri-side change needed is adding `.current_dir()` when spawning the CUDA backend:
+
+```rust
+let cuda_dir = data_dir.join("backends/cuda");
+let exe_path = cuda_dir.join("voicebox-server-cuda.exe");
+
+let mut cmd = app.shell().command(exe_path.to_str().unwrap());
+cmd = cmd.current_dir(&cuda_dir);  // PyInstaller finds all DLLs relative to exe
+cmd = cmd.args(["--data-dir", &data_dir_str, "--port", &port_str, "--parent-pid", &parent_pid_str]);
+```
+
+`.current_dir()` tells the PyInstaller bootloader that everything (DLLs, `nvidia/cublas/lib/`, `_internal/`, torch extensions, etc.) lives relative to the exe. Torch finds the NVIDIA libs exactly as it does in a normal `pip install` or dev environment -- no `LD_LIBRARY_PATH` hacks, no `os.add_dll_directory` gymnastics.
+
+### Community evidence
+
+- Multiple Tauri users run this exact pattern: Nuitka folders (exe + pythonXX.dll + supporting files), multi-file .NET apps, and PyInstaller onedir backends (GitHub issues #5719, discussion #5206).
+- The shell plugin explicitly supports `cwd` in both Rust and JS APIs.
+- No reports of torch/CUDA-specific breakage -- the onedir layout is identical to what PyInstaller produces in normal usage.
+
+### Known gotcha: process termination on Windows
+
+PyInstaller onedir creates a parent bootloader + child Python process on Windows. `child.kill()` only hits the outer process in some cases (Tauri issue #11686). Mitigation: keep a reference to the parent PID or use `taskkill /F /T` for clean shutdown. This is not a blocker -- our existing `--parent-pid` watchdog mechanism in `server.py` already handles orphan cleanup.
+
+## Next Steps
+
+1. Prototype: Build the current CUDA binary with `--onedir` and verify torch CUDA works from the output directory
+2. Measure the size split: how much is NVIDIA libs vs everything else
+3. Design the two-archive download flow and dual version checking
+4. Update `cuda.py` for dual-archive extraction (server core + cuda-libs)
+5. Update `main.rs`: change launch path to `backends/cuda/` dir + add `.current_dir()`
+6. Add `ensure_cuda_structure()` helper in Rust to verify exe + nvidia/ subdirs exist before spawning
+7. Update CI pipeline: `build-cuda-windows` produces two archives instead of split parts
+8. ~~Update `split_binary.py` or replace with archive-based distribution~~ Done: replaced with `package_cuda.py`

justfile

@@ -46,6 +46,8 @@ setup-python:
     {{ pip }} install -r {{ backend_dir }}/requirements.txt
     # Chatterbox pins numpy<1.26 / torch==2.6 which break on Python 3.12+
     {{ pip }} install --no-deps chatterbox-tts
+    # HumeAI TADA pins torch>=2.7,<2.8 which conflicts with our torch>=2.1
+    {{ pip }} install --no-deps hume-tada
     # Apple Silicon: install MLX backend
     if [ "$(uname -m)" = "arm64" ] && [ "$(uname)" = "Darwin" ]; then
         echo "Detected Apple Silicon — installing MLX dependencies..."
@@ -70,10 +72,11 @@ setup-python:
     $hasNvidia = $null -ne (Get-WmiObject Win32_VideoController | Where-Object { $_.Name -match 'NVIDIA' })
     if ($hasNvidia) { \
         Write-Host "NVIDIA GPU detected — installing PyTorch with CUDA support..."; \
-        & "{{ pip }}" install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126; \
+        & "{{ pip }}" install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128; \
     }
     & "{{ pip }}" install -r {{ backend_dir }}/requirements.txt
     & "{{ pip }}" install --no-deps chatterbox-tts
+    & "{{ pip }}" install --no-deps hume-tada
     & "{{ pip }}" install git+https://github.com/QwenLM/Qwen3-TTS.git
     & "{{ pip }}" install pyinstaller ruff pytest pytest-asyncio -q
     Write-Host "Python environment ready."
@@ -205,10 +208,11 @@ build-server-cuda: _ensure-venv
     $env:PATH = "{{ venv_bin }};$env:PATH"; \
     & "{{ python }}" backend/build_binary.py --cuda; \
     if ($LASTEXITCODE -ne 0) { throw "build_binary.py --cuda failed with exit code $LASTEXITCODE" }; \
-    $dest = "$env:APPDATA/com.voicebox.app/backends"; \
+    $dest = "$env:APPDATA/sh.voicebox.app/backends/cuda"; \
+    if (Test-Path $dest) { Remove-Item -Recurse -Force $dest }; \
     New-Item -ItemType Directory -Path $dest -Force | Out-Null; \
-    Copy-Item "backend/dist/voicebox-server-cuda.exe" "$dest/voicebox-server-cuda.exe" -Force; \
-    Write-Host "Copied CUDA binary to $dest"
+    Copy-Item "backend/dist/voicebox-server-cuda/*" $dest -Recurse -Force; \
+    Write-Host "Copied CUDA backend to $dest"
 
 # Build everything locally: CPU server + CUDA server + installable Tauri app
 [windows]

landing/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/landing",
-  "version": "0.2.3",
+  "version": "0.3.1",
   "description": "Landing page for voicebox.sh",
   "scripts": {
     "dev": "bun --bun next dev --turbo",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "voicebox",
-  "version": "0.2.3",
+  "version": "0.3.1",
   "private": true,
   "workspaces": [
     "app",

scripts/package_cuda.py

@@ -0,0 +1,232 @@
+"""
+Package the PyInstaller --onedir CUDA build into two archives.
+
+Takes the PyInstaller --onedir output directory and splits it into:
+  1. voicebox-server-cuda.tar.gz  — server core (exe + non-NVIDIA deps)
+  2. cuda-libs-cu128.tar.gz       — NVIDIA runtime libraries only
+  3. cuda-libs.json                — version manifest for the CUDA libs
+
+Usage:
+    python scripts/package_cuda.py backend/dist/voicebox-server-cuda/
+    python scripts/package_cuda.py backend/dist/voicebox-server-cuda/ --output release-assets/
+    python scripts/package_cuda.py backend/dist/voicebox-server-cuda/ --cuda-libs-version cu128-v1
+"""
+
+import argparse
+import hashlib
+import json
+import sys
+import tarfile
+from pathlib import Path
+
+# DLL name prefixes that identify NVIDIA CUDA runtime libraries.
+# These DLLs may appear in different locations depending on the torch
+# and PyInstaller version:
+#   - nvidia/ subdirectories (older torch with separate nvidia-* packages)
+#   - _internal/torch/lib/ (torch 2.10+ bundles NVIDIA DLLs directly)
+#   - Top-level directory (some PyInstaller versions)
+NVIDIA_DLL_PREFIXES = (
+    "cublas",
+    "cublaslt",
+    "cudart",
+    "cudnn",
+    "cufft",
+    "cufftw",
+    "curand",
+    "cusolver",
+    "cusolvermg",
+    "cusparse",
+    "nvjitlink",
+    "nvrtc",
+    "nccl",
+    "caffe2_nvrtc",
+)
+
+# Files to keep in the server core even if they match NVIDIA prefixes.
+# These are small Python modules or stubs, not the large runtime DLLs.
+NVIDIA_KEEP_IN_CORE = {
+    "torch/cuda/nccl.py",
+    "torch/_inductor/codegen/cuda/cutlass_lib_extensions/cutlass_mock_imports/cuda/cudart.py",
+}
+
+
+def is_nvidia_file(rel_path: str) -> bool:
+    """Check if a relative path belongs to the NVIDIA CUDA libs.
+
+    Identifies large NVIDIA runtime DLLs (.dll/.so) regardless of where
+    PyInstaller placed them. Excludes small Python stubs that happen to
+    share NVIDIA-related names.
+    """
+    rel_lower = rel_path.lower().replace("\\", "/")
+
+    # Never split out Python source files or small stubs
+    if rel_lower in NVIDIA_KEEP_IN_CORE:
+        return False
+
+    # Files under nvidia/ subdirectory tree (older torch layout)
+    if rel_lower.startswith("nvidia/") or "/nvidia/" in rel_lower:
+        # Only DLLs/shared objects — not .py, .dist-info, etc.
+        if rel_lower.endswith((".dll", ".so")):
+            return True
+        # Include entire nvidia/ namespace package tree
+        for part in rel_lower.split("/"):
+            if part == "nvidia":
+                return True
+
+    # NVIDIA DLLs anywhere in the tree (e.g. _internal/torch/lib/cublas64_12.dll)
+    name = rel_lower.rsplit("/", 1)[-1]
+    if name.endswith(".dll") or name.endswith(".so"):
+        name_no_ext = name.rsplit(".", 1)[0]
+        for prefix in NVIDIA_DLL_PREFIXES:
+            if name_no_ext.startswith(prefix):
+                return True
+
+    return False
+
+
+def sha256_file(path: Path) -> str:
+    """Compute SHA-256 hex digest of a file."""
+    h = hashlib.sha256()
+    with open(path, "rb") as f:
+        while True:
+            chunk = f.read(1024 * 1024)
+            if not chunk:
+                break
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def package(
+    onedir_path: Path,
+    output_dir: Path,
+    cuda_libs_version: str,
+    torch_compat: str,
+):
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    # Collect all files in the onedir output, split into core vs nvidia
+    core_files = []
+    nvidia_files = []
+
+    for item in sorted(onedir_path.rglob("*")):
+        if item.is_dir():
+            continue
+        rel = item.relative_to(onedir_path)
+        rel_str = str(rel)
+        if is_nvidia_file(rel_str):
+            nvidia_files.append((rel_str, item))
+        else:
+            core_files.append((rel_str, item))
+
+    core_size = sum(f.stat().st_size for _, f in core_files)
+    nvidia_size = sum(f.stat().st_size for _, f in nvidia_files)
+
+    print(f"Input directory: {onedir_path}")
+    print(f"Core files:  {len(core_files)} ({core_size / (1024**2):.1f} MB)")
+    print(f"NVIDIA files: {len(nvidia_files)} ({nvidia_size / (1024**2):.1f} MB)")
+
+    if not nvidia_files:
+        print(
+            f"ERROR: No NVIDIA files found in {onedir_path}. "
+            "Refusing to create an empty CUDA libs archive.",
+            file=sys.stderr,
+        )
+        print(
+            "Make sure you built with --cuda and the NVIDIA packages are present.",
+            file=sys.stderr,
+        )
+        sys.exit(1)
+
+    # Create server core archive
+    # Files are stored relative to the archive root (no parent directory prefix)
+    # so extracting to backends/cuda/ puts everything at the right level.
+    server_archive = output_dir / "voicebox-server-cuda.tar.gz"
+    print(f"\nCreating server core archive: {server_archive.name}")
+    with tarfile.open(server_archive, "w:gz") as tar:
+        for rel_str, full_path in core_files:
+            tar.add(full_path, arcname=rel_str)
+    server_sha = sha256_file(server_archive)
+    (output_dir / "voicebox-server-cuda.tar.gz.sha256").write_text(
+        f"{server_sha}  voicebox-server-cuda.tar.gz\n"
+    )
+    print(f"  Size: {server_archive.stat().st_size / (1024**2):.1f} MB")
+    print(f"  SHA-256: {server_sha[:16]}...")
+
+    # Create CUDA libs archive
+    cuda_libs_archive = output_dir / f"cuda-libs-{cuda_libs_version}.tar.gz"
+    print(f"\nCreating CUDA libs archive: {cuda_libs_archive.name}")
+    with tarfile.open(cuda_libs_archive, "w:gz") as tar:
+        for rel_str, full_path in nvidia_files:
+            tar.add(full_path, arcname=rel_str)
+    cuda_sha = sha256_file(cuda_libs_archive)
+    (output_dir / f"cuda-libs-{cuda_libs_version}.tar.gz.sha256").write_text(
+        f"{cuda_sha}  cuda-libs-{cuda_libs_version}.tar.gz\n"
+    )
+    print(f"  Size: {cuda_libs_archive.stat().st_size / (1024**2):.1f} MB")
+    print(f"  SHA-256: {cuda_sha[:16]}...")
+
+    # Write cuda-libs.json manifest
+    manifest = {
+        "version": cuda_libs_version,
+        "torch_compat": torch_compat,
+        "archive": cuda_libs_archive.name,
+        "sha256": cuda_sha,
+    }
+    manifest_path = output_dir / "cuda-libs.json"
+    manifest_path.write_text(json.dumps(manifest, indent=2) + "\n")
+    print(f"\nManifest: {manifest_path.name}")
+    print(json.dumps(manifest, indent=2))
+
+    # Summary
+    total_input = core_size + nvidia_size
+    total_output = server_archive.stat().st_size + cuda_libs_archive.stat().st_size
+    print(f"\nTotal input:  {total_input / (1024**3):.2f} GB")
+    print(f"Total output: {total_output / (1024**3):.2f} GB (compressed)")
+    print(
+        f"Server core:  {server_archive.stat().st_size / (1024**2):.1f} MB (redownloaded on app update)"
+    )
+    print(
+        f"CUDA libs:    {cuda_libs_archive.stat().st_size / (1024**2):.1f} MB (cached until CUDA toolkit bump)"
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Package PyInstaller --onedir CUDA build into server + CUDA libs archives"
+    )
+    parser.add_argument(
+        "input",
+        type=Path,
+        help="Path to PyInstaller --onedir output directory (e.g. backend/dist/voicebox-server-cuda/)",
+    )
+    parser.add_argument(
+        "--output",
+        type=Path,
+        default=None,
+        help="Output directory for archives (default: same as input parent)",
+    )
+    parser.add_argument(
+        "--cuda-libs-version",
+        type=str,
+        default="cu128-v1",
+        help="Version string for the CUDA libs archive (default: cu128-v1)",
+    )
+    parser.add_argument(
+        "--torch-compat",
+        type=str,
+        default=">=2.7.0,<2.11.0",
+        help="Torch version compatibility range (default: >=2.6.0,<2.11.0)",
+    )
+    args = parser.parse_args()
+
+    if not args.input.is_dir():
+        print(f"Error: {args.input} is not a directory", file=sys.stderr)
+        print("Expected a PyInstaller --onedir output directory.", file=sys.stderr)
+        sys.exit(1)
+
+    output_dir = args.output or args.input.parent
+    package(args.input, output_dir, args.cuda_libs_version, args.torch_compat)
+
+
+if __name__ == "__main__":
+    main()

scripts/split_binary.py

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

tauri/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/tauri",
   "private": true,
-  "version": "0.2.3",
+  "version": "0.3.1",
   "type": "module",
   "scripts": {
     "dev": "vite",

tauri/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "voicebox"
-version = "0.2.3"
+version = "0.3.1"
 description = "A production-quality desktop app for Qwen3-TTS voice cloning and generation"
 authors = ["you"]
 license = ""

tauri/src-tauri/build.rs

@@ -23,15 +23,15 @@ fn main() {
         }
     }
 
+    let project_root = env!("CARGO_MANIFEST_DIR");
+    let gen_dir = format!("{}/gen", project_root);
+    std::fs::create_dir_all(&gen_dir).expect("Failed to create gen directory");
+
     // Compile macOS Liquid Glass icon
     #[cfg(target_os = "macos")]
     {
-        let project_root = env!("CARGO_MANIFEST_DIR");
         // voicebox.icon is in tauri/assets/voicebox.icon (one level up from src-tauri)
         let icon_source = format!("{}/../assets/voicebox.icon", project_root);
-        let gen_dir = format!("{}/gen", project_root);
-
-        std::fs::create_dir_all(&gen_dir).expect("Failed to create gen directory");
 
         if std::path::Path::new(&icon_source).exists() {
             println!("cargo:rerun-if-changed={}", icon_source);
@@ -76,6 +76,71 @@ fn main() {
                     panic!("Icon compilation failed");
                 }
             }
+
+            // Generate voicebox.icns from the source PNG via sips + iconutil
+            let icns_path = format!("{}/voicebox.icns", gen_dir);
+            if !std::path::Path::new(&icns_path).exists() {
+                let source_png = format!("{}/Assets/Voicebox.png", icon_source);
+                if std::path::Path::new(&source_png).exists() {
+                    let iconset_dir = format!("{}/voicebox.iconset", gen_dir);
+                    std::fs::create_dir_all(&iconset_dir).ok();
+
+                    let sizes: &[(u32, &str)] = &[
+                        (16, "icon_16x16.png"),
+                        (32, "[email protected]"),
+                        (32, "icon_32x32.png"),
+                        (64, "[email protected]"),
+                        (128, "icon_128x128.png"),
+                        (256, "[email protected]"),
+                        (256, "icon_256x256.png"),
+                        (512, "[email protected]"),
+                        (512, "icon_512x512.png"),
+                        (1024, "[email protected]"),
+                    ];
+
+                    for (size, name) in sizes {
+                        let dest = format!("{}/{}", iconset_dir, name);
+                        let status = Command::new("sips")
+                            .args([
+                                "-z",
+                                &size.to_string(),
+                                &size.to_string(),
+                                &source_png,
+                                "--out",
+                                &dest,
+                            ])
+                            .output();
+                        if let Ok(out) = status {
+                            if !out.status.success() {
+                                eprintln!(
+                                    "sips failed for {}: {}",
+                                    name,
+                                    String::from_utf8_lossy(&out.stderr)
+                                );
+                            }
+                        }
+                    }
+
+                    let iconutil_output = Command::new("iconutil")
+                        .args(["-c", "icns", "-o", &icns_path, &iconset_dir])
+                        .output();
+
+                    match iconutil_output {
+                        Ok(out) if out.status.success() => {
+                            println!("Generated voicebox.icns");
+                        }
+                        Ok(out) => {
+                            eprintln!("iconutil failed: {}", String::from_utf8_lossy(&out.stderr));
+                        }
+                        Err(e) => {
+                            eprintln!("Failed to run iconutil: {}", e);
+                        }
+                    }
+
+                    // Clean up iconset directory
+                    std::fs::remove_dir_all(&iconset_dir).ok();
+                }
+            }
         } else {
             println!(
                 "cargo:warning=Icon source not found at {}, skipping icon compilation",
@@ -84,5 +149,18 @@ fn main() {
         }
     }
 
+    // Ensure all resource files exist so Tauri's bundler doesn't fail.
+    // On non-macOS these are always stubs. On macOS, actool may not produce
+    // Assets.car if the Xcode version doesn't support the .icon format.
+    {
+        let required = ["Assets.car", "voicebox.icns", "partial.plist"];
+        for name in required {
+            let path = format!("{}/{}", gen_dir, name);
+            if !std::path::Path::new(&path).exists() {
+                std::fs::write(&path, b"").ok();
+            }
+        }
+    }
+
     tauri_build::build()
 }

tauri/src-tauri/src/main.rs

@@ -197,22 +197,24 @@ async fn start_server(
     println!("Data directory: {:?}", data_dir);
     println!("Remote mode: {}", remote.unwrap_or(false));
 
-    // Check for CUDA backend binary in data directory
+    // Check for CUDA backend in data directory (onedir layout: backends/cuda/)
     let cuda_binary = {
-        let backends_dir = data_dir.join("backends");
+        let cuda_dir = data_dir.join("backends").join("cuda");
         let cuda_name = if cfg!(windows) {
             "voicebox-server-cuda.exe"
         } else {
             "voicebox-server-cuda"
         };
-        let path = backends_dir.join(cuda_name);
-        if path.exists() {
-            println!("Found CUDA backend binary at {:?}", path);
+        let exe_path = cuda_dir.join(cuda_name);
+        if exe_path.exists() {
+            println!("Found CUDA backend at {:?}", cuda_dir);
 
-            // Version check: run --version and compare to app version
+            // Version check: run --version from the onedir directory so
+            // PyInstaller can find its support files for the fast --version path
             let app_version = app.config().version.clone().unwrap_or_default();
-            let version_ok = match std::process::Command::new(&path)
+            let version_ok = match std::process::Command::new(&exe_path)
                 .arg("--version")
+                .current_dir(&cuda_dir)
                 .output()
             {
                 Ok(output) => {
@@ -237,7 +239,7 @@ async fn start_server(
             };
 
             if version_ok {
-                Some(path)
+                Some(exe_path)
             } else {
                 None
             }
@@ -300,10 +302,14 @@ async fn start_server(
         println!("Custom models directory: {}", dir);
     }
 
-    // If CUDA binary exists, launch it directly instead of the bundled sidecar
+    // If CUDA binary exists, launch it from the onedir directory.
+    // .current_dir() is critical: PyInstaller onedir expects all DLLs and
+    // support files (nvidia/, _internal/, etc.) relative to the exe.
     let spawn_result = if let Some(ref cuda_path) = cuda_binary {
-        println!("Launching CUDA backend: {:?}", cuda_path);
+        let cuda_dir = cuda_path.parent().unwrap();
+        println!("Launching CUDA backend: {:?} (cwd: {:?})", cuda_path, cuda_dir);
         let mut cmd = app.shell().command(cuda_path.to_str().unwrap());
+        cmd = cmd.current_dir(cuda_dir);
         cmd = cmd.args(["--data-dir", &data_dir_str, "--port", &port_str, "--parent-pid", &parent_pid_str]);
         if is_remote {
             cmd = cmd.args(["--host", "0.0.0.0"]);

tauri/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Voicebox",
-  "version": "0.2.3",
+  "version": "0.3.1",
   "identifier": "sh.voicebox.app",
   "build": {
     "beforeDevCommand": "bun run dev",

web/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/web",
   "private": true,
-  "version": "0.2.3",
+  "version": "0.3.1",
   "type": "module",
   "scripts": {
     "dev": "vite",