The previous fix (#481) capped transformers at 4.57.6 in requirements-mlx.txt,
but pip's clean resolver in CI can't satisfy that alongside mlx-audio>=0.3.1
(declares `transformers==5.0.0rc3` or `>=5.0.0`) — it backtracks through every
transformers and tokenizers version and exits with `ResolutionImpossible`.
The dev install worked only because mlx-audio 0.4.1 was already present, so
pip never tried to re-resolve.
mlx-audio 0.4.1 + mlx-lm 0.31.1 both declare transformers>=5.x but the API
surface we actually use works fine on 4.57.x in practice (verified across all
engines in dev). Install both --no-deps to bypass the resolver; transitive
runtime deps (huggingface_hub, librosa, numpy, numba, pyloudnorm, etc.) are
already pulled in by requirements.txt.
Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>
mlx-audio depends on `transformers` with no upper bound. Installing
requirements-mlx.txt after requirements.txt lets pip upgrade transformers
past the 4.57.x cap to 5.x, which breaks three engines in the frozen MLX
bundle:
- qwen-custom-voice: `check_model_inputs` was rewritten to take `func` as
positional, so `@check_model_inputs()` factory calls fail with
`TypeError: missing 1 required positional argument: 'func'`
- tada-1b: `PretrainedConfig.__init_subclass__` now applies `@dataclass`,
which rejects tada's `strides: list = []` mutable default
- luxtts: Whisper init hits `AssertionError` in `torch._refs.normal_`
Restating the same constraint here keeps mlx-audio's transformers
dependency from quietly winning the resolver.
Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>
When migrating model path with an empty cache, backend returned early without emitting migration completion SSE, causing frontend overlay to hang. This patch emits complete status for empty migrations.
* fix(build): bundle kokoro source files for transformers runtime introspection
transformers opens .py source files at runtime to check attention/MoE
implementation via regex (e.g. _can_set_attn_implementation). PyInstaller's
--hidden-import only bundles .pyc bytecode, so kokoro/modules.py was missing
from the bundle causing a FileNotFoundError on Kokoro model load.
Switch from individual --hidden-import entries to --collect-all kokoro in both
build_binary.py and voicebox-server.spec. The kokoro package is 172K so no
meaningful bundle size impact.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* fix(build): use SPECPATH for runtime hook instead of hardcoded absolute path
The linter expanded runtime_hooks=[] to an absolute /Users/... path which
would break CI and other dev machines. Use os.path.join(SPECPATH, ...) to
mirror the relative approach in build_binary.py.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* fix(build): runtime hook to work around PyInstaller + Python 3.12 import breakages
Four distinct bundling-specific crashes blocked Kokoro and Qwen CustomVoice
from loading in the frozen binary:
1. torch._dynamo import triggered via class-body decorators
(@torch._dynamo.allow_in_graph on PreTrainedModel,
@torch.compiler.disable in flex_attention) pulls in torch._numpy._ufuncs
which crashes on module load with NameError: name 'name' is not defined.
2. AlbertModel (Kokoro) triggers @auto_docstring -> modeling_auto ->
GenerationMixin -> candidate_generator -> sklearn -> scipy, which hits
the same class of bug in scipy.stats._distn_infrastructure (NameError:
name 'obj' is not defined).
3. AutoModel (Qwen) pulls the same sklearn -> scipy chain directly.
4. librosa (required by most TTS engines) -> scipy.signal -> scipy.stats
hits the _distn_infrastructure crash regardless of the transformers
stubs above.
The root cause of (1) and (4) is that PyInstaller's frozen importer runs
module-level `for X in [<list-comp using dir()>]:` loops with an empty
iterable, leaving the loop variable unbound. Trailing `del obj` / unrelated
references then crash.
Fix: a single runtime hook (pyi_rth_torch_compiler_disable.py) installs:
- sys.modules stubs for torch._dynamo and torch._dynamo.config, plus a
meta-path finder for torch._dynamo.* submodules — voicebox never uses
torch.compile/dynamo for inference, so a permissive no-op stub (callable
as decorator, falsey as predicate, context-manager-safe for
TransformGetItemToIndex) is drop-in safe.
- meta-path finder stubs for transformers.utils.auto_docstring and
transformers.generation.candidate_generator — both import-chain
short-circuits; docstrings and speculative decoding aren't used for TTS.
- meta-path finder for scipy.stats._distn_infrastructure that reads the
real .py source via the wrapped loader's get_source(), replaces the
bundling-broken `del obj` with `globals().pop('obj', None)`, and
compile+exec's the patched source. This keeps the real scipy module
intact so librosa and everything downstream works normally.
Supporting changes:
- backend/pyi_hooks/hook-scipy.stats._distn_infrastructure.py sets
module_collection_mode = "pyz+py" so the .py source is actually in the
bundle for the runtime patcher to read.
- build_binary.py and voicebox-server.spec register the runtime hook and
the new hooks dir.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
* fix(build): force transformers torch<2.6 mask path and bundle spacy_pkuseg
- patch transformers.masking_utils to set _is_torch_greater_or_equal_than_2_6
= False, forcing sdpa_mask_older_torch and avoiding the vmap .item() crash
that breaks Qwen CustomVoice generation (our torch._dynamo stub can't
reproduce TransformGetItemToIndex's graph transform).
- add PyInstaller hook to bundle transformers.masking_utils .py source so the
runtime finder can source-patch it.
- --collect-all spacy_pkuseg so Chatterbox Multilingual can load its Chinese
segmenter (dicts/default.pkl + native .so extensions).
- add per-finder install diagnostics + _HOOK_VERSION marker to make future
bundle-only regressions easier to triage.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(build): pass PyInstaller hook paths relative so .spec is portable
Absolute paths ended up in the auto-regenerated voicebox-server.spec
because build_binary.py prefixed every --runtime-hook and
--additional-hooks-dir with str(backend_dir / ...). That broke builds
on any machine whose checkout wasn't at /Users/jamie/... and anyone
invoking pyinstaller voicebox-server.spec directly.
os.chdir(backend_dir) already runs before PyInstaller (same reason
server.py works as a bare filename), so the backend_dir prefix is
unnecessary. Drop it so the generated spec references pyi_hooks/,
pyi_rth_numpy_compat.py, pyi_rth_torch_compiler_disable.py as repo-
relative paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>
0.3.0 sometimes stored relative media paths with the data-dir name baked in
(e.g. "data/profiles/<uuid>/sample.wav"). resolve_storage_path joined those
directly with _data_dir, producing "<data_dir>/data/profiles/..." — a
spurious double nest that breaks file reads after upgrading to 0.4.0.
The 0.4.0 startup migration didn't catch it because resolve_storage_path
produced the buggy double-nested path, to_storage_path saw "data" at the
first (legitimate) index, and the normalized value matched the stored value
so the row was skipped.
Strip any leading "data/" component before joining. This unblocks runtime
reads and lets _normalize_storage_paths rewrite the affected rows on next
startup — no manual migration needed.
Fixes "No such file or directory: '<data_dir>/data/profiles/...'" and
associated 404s on GET /audio/<id> after upgrading from 0.3.0 to 0.4.0.
Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>
- Show all voice profiles with unsupported ones grayed out (opacity) instead of hidden
- Clicking a grayed-out profile selects it and auto-switches the engine to a compatible one
- Sort supported profiles first, with info tip about compatibility at the bottom
- Scroll to selected profile after engine/sort changes with safe margin
- Fix engine desync on tab navigation by initializing form engine from store
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
When the model wasn't loaded, the app was closed mid-run, or a
generation otherwise errored out, the resulting "Failed" rows
accumulate in history and there was no way to remove them in bulk —
individual delete was the only option.
Adds a header row above the history list (only rendered when at
least one failed generation is present) with a "Clear failed" button
that opens a confirmation dialog, then calls a new
DELETE /history/failed endpoint which sweeps all status='failed'
rows (plus their version files / audio files on disk).
Closesjamiepine/voicebox#410
Co-authored-by: Claude Opus 4.6 <[email protected]>
Applies the compatibility-checker portion of #367. Adds a
check_cuda_compatibility() helper that compares the current device's
compute capability against torch.cuda._get_arch_list() and returns a
human-readable warning if the PyTorch build doesn't support it.
Wired into three places:
• HealthResponse gains a gpu_compatibility_warning field so clients
can surface the issue in the UI
• Startup logs the warning as WARN level
• _get_gpu_status() appends "[UNSUPPORTED - see logs]" to the GPU
label shown in settings
Skipped #367's other half — the switch from stable to nightly cu128
wheels across release.yml, build_binary.py, and justfile. That's
redundant with #401's TORCH_CUDA_ARCH_LIST=...12.0+PTX approach and
would introduce non-deterministic builds from shifting nightly
releases.
Co-Authored-By: nyzxor <[email protected]>
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Applies the cache_dir portion of #218. On Windows local setups, model
assets can split between .hf-cache/hub and .hf-cache/transformers when
Qwen3TTSModel.from_pretrained doesn't explicitly pin the cache root —
speech_tokenizer and preprocessor_config.json then fail to resolve
during load, causing 500s at generation time.
Routes both HF Hub and Transformers through hf_constants.HF_HUB_CACHE.
Skipped the torch_dtype= → dtype= rename from #218: transformers 4.36
(our minimum) doesn't accept the dtype alias, only 4.46+. Once we bump
the minimum we can make that change.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
Follow-up to #402. The sentinel is only removed inside the grace-period
"sentinel found" branch. When the HTTP /watchdog/disable request wins
the race (normal case on macOS/Linux, occasional on Windows), the
_watchdog_disabled=True check returns first and the sentinel is left on
disk indefinitely.
If a later session spawns a fresh server and the user exits without
"keep running", the new watchdog would find that stale sentinel during
its grace period and keep the server alive against user intent.
Wipe any pre-existing sentinel when the watchdog starts so only signals
written during this session's lifetime can influence grace-period
decisions.
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
The HTTP /watchdog/disable request races with process exit on Windows,
causing the watchdog to kill the server before the request arrives.
Added a .keep-running sentinel file as a reliable fallback:
- Tauri writes the file to data_dir before sending the HTTP request
- The watchdog checks for it during the grace period after detecting
parent death
- The file is removed after being read to avoid stale state
This approach works regardless of HTTP timing because file writes
complete synchronously before the Tauri process exits.
Fixes#372
Co-authored-by: Matt Van Horn <[email protected]>
* fix: add asyncio.Lock to prevent concurrent CUDA downloads
The startup auto-update task and the manual download endpoint can both
invoke download_cuda_binary() concurrently. Without mutual exclusion,
both coroutines write to the same temp file path, corrupting the
download. The progress-manager status check is a TOCTOU race because
the status is not set until after several synchronous checks complete.
Add a module-level asyncio.Lock acquired at the top of
download_cuda_binary() so only one download can proceed at a time.
* fix: fast-reject duplicate CUDA download when lock is held
Address CodeRabbit review feedback: check _download_lock.locked()
before awaiting the lock so concurrent callers return immediately
instead of queueing behind the first download. This prevents the
route handler from returning "started" to multiple callers when only
one download actually proceeds.
Two changes to address the race condition causing "Failed to split clip":
Backend (stories.py): Added with_for_update() to the item query in
split_story_item so concurrent requests for the same clip are
serialized via a row lock instead of racing.
Frontend (StoryTrackEditor.tsx): Guard handleSplit with
splitItem.isPending to prevent rapid double-clicks from firing
multiple mutations before the first completes.
Fixes#366
Co-authored-by: Matt Van Horn <[email protected]>
* fix(history): populate status/error/engine/model_size/is_favorited from DB
GET /history/{generation_id} was constructing HistoryResponse without
passing status, error, engine, model_size, or is_favorited from the
DB row. Since HistoryResponse.status defaults to "completed" in the
Pydantic model (models.py:141), this endpoint returned
status="completed" for every generation regardless of the actual DB
state — including jobs still in "loading_model" or "generating", and
even "failed" jobs.
This breaks any client polling /history/{id} for job completion:
the API lies about the status, so the only trustworthy success
signal becomes `audio_path` being non-empty. All other fields left
at their model defaults were similarly masked.
Fix: pass all fields through from the DB row, matching the pattern
used elsewhere in the codebase. The DB model (Generation in
database/models.py) already has all these columns.
* fix(history): apply NULL fallbacks to match list endpoint
Align the defensive mappings with services/history.py:206-223 so
both the single-item and list history endpoints handle legacy rows
with NULL status/engine/is_favorited identically. Without this,
HistoryResponse's non-Optional str/bool fields would raise a
pydantic ValidationError (500) on any row where these columns are
NULL — possible from direct SQL updates or past migrations.
Addresses review feedback on PR #394.
---------
Co-authored-by: malletfils <[email protected]>
Two small safety improvements:
1. Voice prompt cache (cache.py): add weights_only=True to torch.load()
so cached .prompt files are loaded using the safe unpickler instead of
the unrestricted pickle deserializer. This follows the PyTorch 2.6+
best practice of opting in to safe loading for all torch.load() calls.
2. SPA catch-all (app.py): replace str.startswith() path guard with
Path.is_relative_to(). The string prefix check passes for sibling
paths like /app/frontend_evil/ that share the /app/frontend prefix.
is_relative_to() correctly tests directory containment.
Follow-up to #361. The original fallback silently mapped unknown numpy
dtypes to torch.float32, which would reinterpret the memcpy'd bytes in
the wrong dtype and corrupt data (e.g. fp16 tensors from some TTS
engines) rather than erroring loudly.
- Hoist dtype_map out of the inner function so it's built once
- Add float16, complex64, complex128 mappings
- Raise TypeError on unknown dtype instead of silent float32 fallback
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
torch is compiled against numpy 1.x. numpy 2.x changed the ABI version
returned by PyArray_GetNDArrayCVersion() (0x01000009 → 0x02000000), so
torch's is_numpy_available() always returns False and torch.from_numpy()
raises RuntimeError. This causes TTS generation to fail with:
ValueError: Unable to create tensor, you should probably activate
padding with 'padding=True'
Two fixes:
1. Pin numpy<2.0 in requirements.txt so new builds bundle a compatible
numpy version. (The existing comment already flagged this intention
but the upper bound was never added.)
2. Add a PyInstaller runtime hook (pyi_rth_numpy_compat.py) that installs
a ctypes memmove fallback for torch.from_numpy() at startup. Runtime
hooks run after FrozenImporter is registered so frozen torch is
importable. The fallback catches RuntimeError from the C-level ABI
check and copies the numpy array into a new tensor via raw memory copy,
bypassing the check entirely. This is a belt-and-suspenders fix that
works regardless of the bundled numpy version.
Co-authored-by: aimaaaimaa <[email protected]>
Co-authored-by: Claude Sonnet 4.6 <[email protected]>
Add Kokoro-82M as a new TTS engine — 82M params, CPU realtime, 8 languages,
Apache 2.0. Unlike cloning engines, Kokoro uses pre-built voice styles, which
required a new profile type system to support non-cloning engines cleanly.
Kokoro engine:
- New kokoro_backend.py implementing TTSBackend protocol
- 50 built-in voices across en/es/fr/hi/it/pt/ja/zh
- KPipeline API with language-aware G2P routing via misaki
- PyInstaller bundling for misaki, language_tags, espeakng_loader, en_core_web_sm
Voice profile type system:
- New voice_type column: 'cloned' | 'preset' | 'designed' (future)
- Preset profiles store engine + voice ID instead of audio samples
- default_engine field on profiles — auto-selects engine on profile pick
- Create Voice dialog: toggle between 'Clone from audio' and 'Built-in voice'
- Edit dialog shows preset voice info instead of sample list for preset profiles
- Engine selector locks to preset engine when preset profile is selected
- Profile grid filters by engine — shows Kokoro voices when Kokoro selected
- Custom empty state when no preset profiles exist for selected engine
Bug fixes:
- Fix relative audio paths in DB causing 404s in production builds
- config.set_data_dir() now resolves to absolute paths
- Startup migration converts existing relative paths to absolute
Also updates PROJECT_STATUS.md and tts-engines.mdx developer guide.
Address CodeRabbit review feedback and user-reported GPU acceleration failure:
- Use shared manual_seed() in chatterbox, chatterbox_turbo, and luxtts
backends so XPU (and future accelerators) get proper device seeding
- Add XPU branch to _get_gpu_status() so startup log reports Intel Arc
GPUs instead of 'None (CPU only)'
- Add XPU VRAM reporting and correct backend_variant fallback in the
/health endpoint
- Switch justfile GPU detection from Get-WmiObject to Get-CimInstance,
simplify the Arc regex to match 'Arc' (not 'Intel.*Arc'), log
detected GPUs, and print manual install instructions on miss
Resolves the root cause where IPEX was silently not installed due to
WMI detection failure, causing CPU-only fallback on Intel Arc systems.
Auto-detect Intel Arc GPUs during Windows setup and install PyTorch
with XPU support + intel-extension-for-pytorch. Enable allow_xpu=True
on all TTS backends (Chatterbox, Chatterbox Turbo, Hume TADA, LuxTTS)
that previously only supported CUDA. Add shared empty_device_cache()
and manual_seed() helpers in base.py to handle XPU memory management
and reproducible seeding alongside CUDA.
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
Upgrade CUDA toolkit from 12.6 (cu126) to 12.8 (cu128) for proper
RTX 50-series (Blackwell) GPU support. Users with RTX 5070/5080/5090
were reporting CUDA detection failures with cu126.
Also fix the GPU Acceleration settings panel where the 'Switch to CPU
Backend' button was unreachable — it was inside a conditional block
that required !isCurrentlyCuda, making it impossible to switch back
to CPU once running on CUDA.
Closes#315
Replace --collect-submodules + --collect-data with --collect-all for
qwen_tts. The qwen_tts runtime expects physical .py source files
(e.g. modeling_qwen3_tts.py) under _MEIPASS, which only --collect-all
provides. This is the same pattern used for inflect/typeguard.
Fixes#212
- Wrap download/verify/extract in try/finally so .download-*.tmp is
always deleted, even on mid-download or extraction failures
- Fix justfile build-server-cuda to use sh.voicebox.app (production path)
Switch CUDA builds from PyInstaller --onefile to --onedir and split the
output into two separately versioned archives:
1. Server core (~200-400MB) — versioned with the app, redownloaded on
every app update
2. CUDA libs (~2GB) — versioned independently (cu126-v1), only
redownloaded when the CUDA toolkit or torch version changes
This eliminates the ~2.4GB full redownload on every version bump.
After initial setup, most app updates only need ~200-400MB.
Closes#297
Remove @torch.jit.script from the DAC shim's snake() function —
TorchScript calls inspect.getsource() which fails in PyInstaller
binaries (no .py source files).
Update all user-facing docs: 4 → 5 TTS engines, add TADA row to
every engine comparison table, mark TADA as Shipped in the upcoming
engines list, update architecture diagrams and tech stack tables.
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.
