Replace the monkey-patch on AutoTokenizer.from_pretrained (which broke
the classmethod descriptor and caused 'Tokenizer not loaded' errors
when loading Qwen after TADA) with two targeted config patches:
- Set AlignerConfig.tokenizer_name to the local ungated tokenizer path
- Pre-load TadaConfig, inject tokenizer_name, pass config= to from_pretrained
No global state is modified; other engines are unaffected.
torchaudio 2.10+ switched its default audio loading backend to
torchcodec, which isn't installed. Replace torchaudio.load() with
soundfile.read() in create_voice_prompt(). TADA's internal use of
torchaudio.functional.resample() is unaffected (pure PyTorch math,
no torchcodec dependency).
TADA hardcodes 'meta-llama/Llama-3.2-1B' as its tokenizer source in
both the Aligner and TadaForCausalLM.from_pretrained(). That repo is
gated and requires accepting Meta's license on HuggingFace.
Monkey-patch AutoTokenizer.from_pretrained during model loading to
redirect Llama tokenizer requests to 'unsloth/Llama-3.2-1B', an
ungated mirror with identical tokenizer files. The patch is scoped
to model loading only and restored immediately after.
The real descript-audio-codec package pulls in descript-audiotools,
which transitively requires onnx, tensorboard, protobuf, matplotlib,
pystoi, and other heavy dependencies. onnx fails to build from source
on macOS due to CMake version incompatibility.
TADA only uses Snake1d (a 7-line PyTorch module) from DAC. This commit
adds a shim in backend/utils/dac_shim.py that registers fake dac.*
modules in sys.modules with just the Snake1d class, completely
eliminating the DAC/audiotools dependency chain.
Integrates HumeAI's TADA (Text-Acoustic Dual Alignment) speech-language
model as a new TTS engine. TADA uses a novel 1:1 token-audio alignment
that produces coherent speech over long sequences (700s+).
Two model variants:
- tada-1b: English-only, ~4GB, built on Llama 3.2 1B
- tada-3b-ml: 10 languages, ~8GB, built on Llama 3.2 3B
Backend uses the Encoder for voice prompt encoding with caching, and
TadaForCausalLM with flow-matching diffusion for generation. Supports
bf16 inference on CUDA, forces CPU on macOS (MPS compatibility).
Installed with --no-deps due to torch>=2.7 pin conflict; descript-audio-codec
and torchaudio added as explicit sub-dependencies.
- Use DB COUNT query instead of list length for take-N label to avoid
TOCTOU race between list_versions and create_version
- Add focus:bg-muted to SelectTrigger for keyboard focus visibility
- Force tqdm disable=False in TrackedTqdm so byte progress works in prod
(huggingface_hub disables tqdm based on logger level, which prevents
self.n from updating — our progress tracking needs the counter even
though we don't render to terminal)
- Harden devnull redirect to test writability, not just None check
- Add full traceback logging to all backend error handlers
- Add chatterbox/luxtts/zipvoice hidden imports and metadata to spec
Apply local-var-then-assign pattern to chatterbox_backend.py (multilingual)
to match the turbo backend. Also use _current_model_size fallback in
unload, delete, and status endpoints for consistent Qwen model size checks.
The actual dtype mismatch was in S3Tokenizer.log_mel_spectrogram, not
VoiceEncoder.forward. librosa.load returns float64 numpy, which
torch.from_numpy preserves as double. The STFT output (double) then
hits _mel_filters (float32) in a matmul at s3tokenizer.py:163.
Now patching both entry points after model load:
1. S3Tokenizer.log_mel_spectrogram — cast audio to float32 before STFT
2. VoiceEncoder.forward — cast mels to float32 before LSTM
Remove debug traceback logging (no longer needed).
The previous approach of patching librosa.load didn't work because
melspectrogram itself performs float64 math (numpy dot, signal.lfilter)
regardless of input dtype. The actual mismatch happens when pack()
creates a float64 tensor from the mel arrays and passes it into the
float32 LSTM weights in VoiceEncoder.forward().
Fix by monkey-patching VoiceEncoder.forward() to call mels.float()
before the LSTM, ensuring the input always matches the model dtype.
The upstream VoiceEncoder's melspectrogram only casts to float32 when
hp.normalized_mels is True (it defaults to False), so librosa's float64
output flows through as double tensors into float32 model weights,
causing 'expected m1 and m2 to have the same dtype, but got: float !=
double'. Fix by monkey-patching prepare_conditionals in both Chatterbox
and Chatterbox Turbo backends to ensure librosa.load returns float32.
Both PyTorch and MLX backends silently dropped the language parameter —
it was accepted by generate() but never forwarded to the underlying
Qwen3-TTS model, causing it to default to auto-detection which
frequently confuses similar languages (e.g. Portuguese for Spanish).
- Add LANGUAGE_CODE_TO_NAME mapping (ISO 639-1 to full name) to both backends
- PyTorch: pass language= to generate_voice_clone()
- MLX: pass lang_code= to all 4 model.generate() call sites
- Frontend: auto-sync generation form language with selected voice profile
Closes#97
Cherry-picked and adapted from PR #89 and #214:
- Linux audio capture via PulseAudio/PipeWire monitor sources (cpal)
- AMD ROCm GPU support: HSA_OVERRIDE_GFX_VERSION env var, ROCm detection
- Whisper Turbo model (openai/whisper-large-v3-turbo) in all endpoints
- Cleaner Whisper language handling via generate_kwargs
- tauri::async_runtime::spawn fix to prevent panic on app shutdown
- Enable Linux (ubuntu-22.04) in release CI matrix
- New ChatterboxTurboTTSBackend wrapping ChatterboxTurboTTS (ResembleAI/chatterbox-turbo)
- English-only 350M model with paralinguistic tag support ([laugh], [cough], [chuckle])
- Bypasses upstream token=True bug by calling snapshot_download(token=None) + from_local()
- Same CPU-on-macOS forcing and torch.load monkey-patching as multilingual backend
- Full engine integration: generate, stream, model status/download/delete endpoints
- Language dropdown now shows only languages supported by the selected engine
- Per-engine language maps: Qwen (10), LuxTTS (en), Chatterbox (23), Turbo (en)
- Auto-switches to English when selecting English-only engines
- Backend language regex expanded to accept all 23 Chatterbox languages
- Add HFProgressTracker to LuxTTS and Chatterbox backends so tqdm-based
file-level download progress reaches the frontend (previously only Qwen
had this, LuxTTS/Chatterbox showed a static spinner)
- Add progress/current/total/filename fields to ActiveDownloadTask so the
/tasks/active polling endpoint carries progress data
- Show inline progress bar + bytes in the model list and detail modal,
poll at 1s during active downloads (5s otherwise)
- Fix GpuAcceleration crash: cudaStatusLoading was referenced before
initialization in its own useQuery declaration
chatterbox-tts 0.1.6 pins numpy<1.26 and torch==2.6 which are
incompatible with Python 3.12+. Install with --no-deps and list
its sub-dependencies explicitly in requirements.txt.
Also removes HFProgressTracker from chatterbox backend to avoid
'generator didn't stop after throw()' errors from tqdm patching.
- New ChatterboxTTSBackend wrapping ChatterboxMultilingualTTS (ResembleAI/chatterbox)
- Supports 23 languages including Hebrew, forces CPU on macOS (MPS issue)
- Monkey-patches torch.load for CPU loading, forces eager attention for compatibility
- trim_tts_output utility cuts trailing silence/hallucination from Chatterbox output
- Full engine integration: /generate, /generate/stream, model status/download/delete
- Hebrew (he) added to supported languages in frontend and backend validation
- Single flat model dropdown extended with Chatterbox option in both generation UIs
- ModelManagement UI groups LuxTTS and Chatterbox under 'Other Voice Models' section
- Add threading lock to get_tts_backend_for_engine() to prevent race
condition where concurrent requests could create duplicate backend
instances (double-checked locking pattern)
- Fix LuxTTS generate: call .detach().cpu() before .numpy() so it
works on GPU/MPS devices, not just CPU
- Store background download tasks in a module-level set to prevent
garbage collection before completion (asyncio.create_task fire-and-
forget pattern)
- Deduplicate cache_key computation in LuxTTS create_voice_prompt
- Prefix unused sr variable with underscore
Introduce LuxTTS (ZipVoice) alongside Qwen TTS, enabling users to choose
between engines at generation time. LuxTTS offers fast, English-focused
voice cloning at 48kHz with ~1GB VRAM.
Backend:
- Add LuxTTSBackend with encode_prompt/generate_speech integration
- Multi-engine registry (get_tts_backend_for_engine) replacing singleton
- Engine-prefixed voice prompt cache keys to avoid collisions
- Engine field on GenerationRequest (default 'qwen' for backward compat)
- Engine dispatch in /generate and /generate/stream endpoints
- LuxTTS in model status, download, and delete maps
Frontend:
- TTS Engine selector dropdown in GenerationForm (Qwen TTS / LuxTTS)
- Conditionally hide Model Size and Delivery Instructions for LuxTTS
- Engine field added to TypeScript types and Zod schema
- LuxTTS section in Model Management page
- Add cancel (X) button on downloading and errored model items
- Add collapsible Problems panel (VS Code-style) showing error details
- Add "Clear All" button to reset all stale download/error state
- Add POST /models/download/cancel endpoint to dismiss individual downloads
- Add POST /tasks/clear endpoint to reset all task and progress state
- Include error messages in /tasks/active response for visibility
- Capture SSE error messages client-side for immediate display
- Fix whisper-large using wrong HF repo (openai/whisper-large → openai/whisper-large-v3)
- Fix Whisper HF repo mapping in both PyTorch and MLX backends
- Shorten error toast to point users to Problems panel instead of wall of text
Implements offline mode patch for API stability issues:
- Add hf_offline_patch.py to monkey-patch huggingface_hub
- Force cache-only lookups before mlx_audio imports
- Create symlink from original Qwen repo to MLX community version
when only MLX version is cached
This fixes:
- Issue #150: Internet required even with cached models
- Issue #151: API crashes when HF network fails
The patch ensures that if models are locally cached, no network
requests are made to HuggingFace during speech generation.
- Updated caching methods in MLX, PyTorch, and backend to ensure models are fully downloaded before being marked as cached.
- Improved progress tracking to filter out non-download progress and provide accurate feedback during model downloads.
- Enhanced HFProgressTracker to skip non-byte progress bars and ensure meaningful progress reporting.
- Refactored progress initialization to provide immediate feedback while fetching metadata from HuggingFace.
- Added error handling and logging for better debugging during cache checks and download processes.
- Rearranged imports for consistency in useModelDownloadToast hook.
- Improved logging in useModelDownloadToast for better debugging during download events.
- Updated progress calculation to handle cases where progress exceeds 100%.
- Enhanced toast notifications to reflect download completion and error states.
- Introduced throttling in ProgressManager to optimize SSE updates and prevent overwhelming clients.
- Added new test scripts for monitoring SSE events during model downloads, ensuring accurate progress reporting.
- Introduced methods to check if models are cached locally in MLX and PyTorch backends.
- Enhanced progress tracking during model loading to filter out non-download progress when models are cached.
- Updated HFProgressTracker to conditionally report progress based on download status.
- Added test scripts for monitoring SSE events during model downloads and verifying progress tracking functionality.
- Improved overall error handling and logging for better debugging during model download processes.
- Updated HistoryTable to implement infinite scrolling for loading history items dynamically.
- Introduced state management for accumulated history and total item count.
- Added Intersection Observer for triggering additional data fetches when scrolling.
- Implemented cache clearing functionality in the backend to manage voice prompt caches effectively.
- Improved loading indicators and user feedback for data fetching states.
- Refactored code for better readability and maintainability.
- Updated README.md to highlight MLX backend performance improvements on Mac with Metal acceleration.
- Refined ProfileCard and ProfileForm components by optimizing imports and improving error handling for avatar uploads.
- Adjusted landing page content to better describe features, including a new multi-voice narrative editor and performance optimizations for different platforms.
- Bumped version to 0.1.11 in Cargo.lock to reflect recent changes.
- Updated hidden imports in build_binary.py to replace 'mlx_audio.asr' with 'mlx_audio.stt'.
- Enhanced model loading logic in MLX and PyTorch backends to ensure proper progress tracking during model downloads.
- Improved error handling and context management for progress tracking in both backends.
- Bumped version to 0.1.10 in Cargo.lock to reflect recent changes.
- Added support for MLX backend on Apple Silicon, enabling optimized performance for TTS and STT tasks.
- Implemented platform detection to dynamically select between MLX and PyTorch based on the runtime environment.
- Updated build process to include MLX-specific dependencies and configurations for macOS.
- Refactored backend code to improve model loading and inference logic, accommodating backend-specific requirements.
- Enhanced documentation to clarify backend selection and performance benefits for different platforms.
- Streamlined installation instructions and troubleshooting guidance for MLX-related issues.
- Added support for MLX backend on Apple Silicon, enabling optimized performance for TTS and STT tasks.
- Updated release workflow to include MLX-specific dependencies and configurations for macOS platforms.
- Refactored backend code to dynamically select between MLX and PyTorch based on the runtime environment.
- Enhanced model loading and inference logic to accommodate backend-specific requirements, including updated model IDs and hidden imports.
- Improved health check and model status reporting to reflect the active backend type.
- Streamlined caching mechanisms to support both backend types, ensuring compatibility and performance.
James Pine