pyproject.toml

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[project]
name = "omnivoice"
version = "0.5.6"
description = "VoiceStudio — a private, local-first studio for voice cloning, speech generation, dubbing, transcription, and audiobooks"
readme = "README.md"
# Free and open-source under the GNU Affero General Public License v3 (see
# LICENSE). A commercial license is available for proprietary/closed-source use
# without AGPL obligations — contact hi@voicestudio.sh. The bundled omnivoice/
# TTS model by Han Zhu remains Apache-2.0 upstream (Apache-2.0 is AGPL-compatible).
license = "AGPL-3.0-only"
requires-python = ">=3.11"
authors = [{name = "Debpalash"}, {name = "Han Zhu"}]
keywords = [
    "tts",
    "text-to-speech",
    "speech-synthesis",
    "zero-shot",
    "multilingual",
    "diffusion",
    "voice-cloning",
]
classifiers = [
    "Intended Audience :: Science/Research",
    "Intended Audience :: Developers",

    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: Multimedia :: Sound/Audio :: Speech",

    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
]
dependencies = [
    "torch>=2.4",
    "torchaudio>=2.4",
    "torchvision>=0.19",
    "transformers>=5.10.0",
    "accelerate",
    "pydub",
    "gradio>=6.15.1",
    "tensorboardX",
    "webdataset",
    "numpy",
    "soundfile",
    # whisperx / faster-whisper import `pkg_resources` at runtime. setuptools
    # 80+ DROPPED the bundled pkg_resources, so an unpinned ">=75" now resolves
    # to a version WITHOUT it → "No module named 'pkg_resources'", which both
    # breaks WhisperX transcription and makes its is_available() report false
    # ("No ASR backend is ready"). Cap below 80 so pkg_resources stays present
    # (issue #224; also #58). Revisit when whisperx/faster-whisper drop the
    # pkg_resources dependency.
    "setuptools>=75,<80",
    "psutil>=7.2.2",
    # Pinned to 3.x — pyannote 4.x removed `use_auth_token` from `Inference`
    # which whisperx 3.4.2 still passes, blowing up `whisperx.load_model()`
    # with TypeError. whisperx tests against pyannote 3.3.2+, so we track
    # that range and revisit when whisperx releases a 4-compatible build.
    "pyannote-audio>=3.3.2,<4.0",
    "imageio-ffmpeg>=0.6.0",
    # Directly used by video_context; 12.1 adds get_flattened_data(), the
    # replacement for getdata() ahead of its Pillow 14 removal.
    "pillow>=12.1.0",
    # 0.9.21+ Linux x86_64 wheels are built with -march=native
    # (spotify/pedalboard#454), so importing pedalboard SIGILLs on CPUs
    # that lack the builder's SIMD set. 0.9.20 is the last known-good
    # wheel. Cap below 0.9.21; revisit when upstream ships portable
    # wheels (USE_PORTABLE_SIMD / pedalboard PR #466). VoiceStudio #2052.
    "pedalboard>=0.9.14,<0.9.21",
    # Primary ASR — cross-platform, CTranslate2-based under the hood. WhisperX
    # adds wav2vec2 forced alignment (±10-30 ms word timing vs Whisper's own
    # ±100-300 ms) which directly improves lip-sync on the dub pipeline.
    # Pulls `faster-whisper` transitively, so a WhisperX install also provides
    # the plain faster-whisper backend as a fallback for rare-language audio
    # where no wav2vec2 alignment model exists.
    "whisperx>=3.1.0",
    "faster-whisper>=1.0.0",
    # Apple Silicon-only speedup; skipped everywhere else so `uv sync` can
    # succeed on Linux/Windows/mac-Intel (no mlx wheels exist for those).
    "mlx-whisper>=0.2.1 ; sys_platform == 'darwin' and platform_machine == 'arm64'",
    # Apple Silicon-only rich TTS library — 14+ engines (Kokoro, CSM, Dia,
    # Qwen3-TTS, Chatterbox, MeloTTS, OuteTTS, Spark, Higgs-Audio, Voxtral,
    # …). Gives mac-ARM users a broad engine picker in Settings. Also gated
    # by platform markers because it depends on mlx (Apple Silicon only).
    "mlx-audio>=0.3.0 ; sys_platform == 'darwin' and platform_machine == 'arm64'",
    # Apple Silicon-only Parakeet TDT v3 ASR (mlx-community/parakeet-tdt-0.6b-v3
    # via MLX). Gives mac-ARM the Parakeet tier CUDA/CPU users get from
    # sherpa/NeMo: 25 European languages, TDT token/word timestamps, ~2 GB
    # unified memory, dictation-grade speed on the GPU. Same platform gate as
    # mlx-whisper/mlx-audio (depends on mlx — no wheels elsewhere).
    "parakeet-mlx>=0.5.2 ; sys_platform == 'darwin' and platform_machine == 'arm64'",
    "demucs>=4.0.1",
    "yt-dlp>=2026.7.4",
    # >=1.16: alembic.ini relies on path_separator=os (new in 1.16.0), which
    # older alembic silently ignores and then colon-splits C:\ paths /
    # space-splits POSIX paths containing spaces.
    "alembic>=1.16",
    # Lightweight English TTS "Turbo" tier — 25-80 MB ONNX model, 8 preset
    # voices (Bella, Jasper, Luna, Bruno, Rosie, Hugo, Kiki, Leo), CPU
    # realtime on any platform. Complements OmniVoice's 2.4 GB multilingual
    # zero-shot clone: when the caller just needs fast English narration with
    # no reference sample, this is ~100× smaller + ~10× faster. Pinned to
    # the exact wheel because the project is in developer preview and the
    # 0.x API is explicitly unstable.
    "kittentts @ https://github.com/KittenML/KittenTTS/releases/download/0.8.1/kittentts-0.8.1-py3-none-any.whl",
    # Invisible audio watermarking — embeds imperceptible neural watermarks
    # in AI-generated speech for provenance detection (SynthID-like).
    # MIT license, ~5ms per segment on CPU, 16-bit message payload.
    "audioseal>=0.1.3",
    # API server — always needed for the Studio UI.
    # 0.137 makes included routers lazy `_IncludedRouter` entries, breaking
    # route-table consumers that require the concrete HTTP/WebSocket routes.
    "fastapi<0.137",
    "scalar-fastapi",
    "uvicorn",
    "python-multipart>=0.0.31",
    # Required by uvicorn for WebSocket support (real-time sidebar events).
    "websockets",
    # Offline translation — listed as builtin in the engine registry so the
    # "Argos (Local, Fast)" option in the Dub tab works out-of-the-box.
    "argostranslate>=1.9.0",
    # Phase 1 AUTH-02: Fernet symmetric encryption + scrypt KDF for the
    # at-rest HF token in the SQLite settings store. Pulled in directly so
    # we don't depend on a transitive arrival via pyannote/huggingface_hub
    # (Assumption A1 in RESEARCH.md was checked at execute-time and proved
    # false — `cryptography` is not on the install path today).
    "cryptography>=41",
    "mcp>=1.28.1,<2",
    # Opt-in product analytics (core/analytics.py). Inert unless the build ships a
    # POSTHOG_PROJECT_TOKEN *and* the user opts in — default OFF. No exception
    # autocapture (it would ship raw tracebacks past core.failure.sanitize()).
    "posthog>=3.7",
    # Fast model downloads (FDL plan). huggingface_hub arrives transitively via
    # transformers, but we pin it directly so the Xet fast-download path can't
    # silently disappear on a resolve, and we pull `hf-xet` explicitly: it is
    # the content-defined-chunking, parallel byte-range downloader that gives
    # IDM/uGet-style speed for Xet-backed repos (the entire current model
    # catalog — FDL spike 2026-06-13; plan removed with .planning/, see git history).
    # hf-xet is 64-bit only (fine for every OmniVoice target). Do NOT add
    # `hf_transfer` — it is deprecated in favour of Xet and breaks progress
    # callbacks (the accurate-progress work in this same plan depends on tqdm).
    "huggingface_hub>=1.7",
    "hf-xet>=1.1",
    # Audiobook PDF ingest (/audiobook/import). Pure-Python, MIT, zero native
    # deps → identical behaviour on macOS/Windows/Linux (default-parity rule).
    # EPUB + plaintext stay stdlib-only; only PDF needs a real parser, and
    # pypdf is the lightest one that ships no C extensions.
    "pypdf>=6.19.0",
    # LLM client for Cinematic dub refinement, glossary auto-extract, and
    # LLM-based translation (services.llm_backend / translator / dub_translate
    # all `from openai import OpenAI`). Talks to any OpenAI-compatible endpoint —
    # OpenAI, Ollama (http://localhost:11434/v1), LM Studio, vLLM — so it's the
    # local-first path too (no key, nothing leaves the machine). Pure-Python, no
    # native deps → identical on macOS/Windows/Linux. Previously undeclared, so
    # `uv sync` never installed it and Cinematic was dead-on-arrival on every
    # source install ("Cinematic needs an LLM" even with Ollama running, because
    # is_available() returned "openai package missing"); the UI's `pip install
    # openai` hint landed in the wrong interpreter on a managed venv.
    "openai>=1.40",
    # Native paid-provider protocols behind the same LLM feature interface.
    "litellm>=1.103.0,<2",
    "httpx>=0.28.1",
    "boto3>=1.40",
    "google-auth>=2.29,<3",
    # sherpa-onnx live-dictation ASR engine (CPU, cross-platform). The thin
    # `sherpa-onnx` wheel declares `sherpa-onnx-core` only in its *wheel*
    # metadata (not the sdist), so uv's resolver does NOT pull it transitively
    # — without core, `import sherpa_onnx` fails at load time (missing
    # libonnxruntime). Pin core EXPLICITLY so the lock captures it and Docker's
    # frozen `uv sync` installs a working engine on every platform.
    "sherpa-onnx>=1.13.3",
    "sherpa-onnx-core>=1.13.3",
    # SOCKS proxy support for httpx (#959). huggingface_hub's get_session()
    # builds an httpx.Client, which raises ImportError AT CONSTRUCTION when
    # ALL_PROXY/HTTPS_PROXY is socks5:// and socksio isn't importable — every
    # model load/download 500'd for SOCKS-proxy users ("Using SOCKS proxy, but
    # the 'socksio' package is not installed"). Same failure shape for the
    # OpenAI SDK's client. Pure-Python, MIT, zero transitive deps, ~13 KB —
    # identical on macOS/Windows/Linux.
    "socksio>=1.0",
    # OS trust store for TLS (#976). Users behind a corporate/antivirus proxy
    # that TLS-inspects traffic get a raw "[SSL: SSLV3_ALERT_HANDSHAKE_FAILURE]"
    # on every model install — the TCP connection succeeds, but the proxy's
    # re-signed certificate is trusted by the OS (Windows CryptoAPI/SChannel)
    # and not by Python's bundled `certifi` CA list. `truststore` patches
    # `ssl.SSLContext` to verify against the OS trust store instead. Pure-
    # Python, MIT, PyPA-maintained, zero transitive deps — same class of fix
    # as socksio above, identical on macOS/Windows/Linux.
    "truststore>=0.9",
    # Numbers→words for the pre-TTS text normalization pass
    # (services/text_normalization.py). Was already installed transitively;
    # promoted to a direct dependency because we now import it ourselves.
    "num2words>=0.5.14",
    "pip>=26.1.2",
    # Direct URL (not [tool.uv.sources]) so EVERY installer sees it — Docker's
    # `uv pip install --system .` reads only project metadata and would try to
    # resolve a bare name from PyPI, where spaCy models don't exist. Same form
    # as kittentts above.
    "en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl",
    # Remote GPU workers (backend/worker/). A core dependency rather than an
    # extra even though the feature is opt-in: an installer user who enables a
    # remote worker in Settings cannot run `uv pip install` inside the
    # installed app, so shipping the transport separately would make the feature
    # source-installs-only — a platform/packaging split the parity rule does
    # not allow. Ships prebuilt wheels for every target we build (cp311
    # macOS arm64/x86_64, Windows x64, manylinux).
    "grpcio>=1.60",
    "protobuf>=5.29.6",
    # Security floors for pyannote-audio's checkpoint loader (GHSA-qqmf-gpg7-g8gw).
    # Declared here, not only in uv.lock, so `uv pip install .` (Docker, Colab)
    # cannot keep a preinstalled 2.6.5 that pyannote's `lightning>=2.0.1` allows.
    "lightning>=2.6.6",
    "pytorch-lightning>=2.6.6",
    "nvidia-ml-py>=13.0",
]

[project.optional-dependencies]

eval = [
    "jiwer==3.1.0",       # WER
    "librosa",             # Audio processing
    "s3prl",               # Speech representation (HuBERT etc.)
    "funasr",              # ASR models
    "zhconv",              # Chinese character normalization
    "zhon",                # Chinese punctuation
    "unidecode",            # Unicode normalization
]
ui = [
    "gradio>=6.15.1",
    "gradio_client",
    "requests",
]
# Phase 3 Plan 03-01 — Supertonic-3 opt-in engine. CPU-only ONNX TTS,
# 31 languages, ~99M params, ~400 MB model on first use. Default
# `uv sync --no-dev` does NOT install this; users opt in with
# `uv sync --extra supertonic` after accepting the OpenRAIL-M model
# license in Settings → Engines.
#
# Publisher verified per Plan 03-01 Task 1 (Package Legitimacy Audit):
#   • PyPI maintainers = Yu Yechan / Juheon Lee / Hyeongju Kim (Supertone Inc.)
#   • Repository = github.com/supertone-inc/supertonic-py
#   • Same publisher ships supertonic-js on npm (same maintainer email)
#   • Wheel inspected: pure-Python, no postinstall scripts, no subprocess/exec
#     at module top level. ``supertonic.config.MODEL_CONFIGS["supertonic-3"]``
#     itself pins the HF model revision by SHA — we re-pin to the same
#     SHA in backend/engines/supertonic3/constants.py for TTS-03.
supertonic = [
    "supertonic==1.3.1",
]
# PocketTTS opt-in engine. Keep the SDK pinned: its model-loading and voice
# APIs are the sidecar wire contract. The model weights remain an explicit
# user download after the Hugging Face access conditions are accepted.
pockettts = [
    # pocket-tts requires torch>=2.5; PyTorch no longer publishes macOS x86_64
    # wheels at those versions. Keep the opt-in extra installable everywhere,
    # with an explicit engine availability reason on Intel Macs.
    "pocket-tts==2.1.0 ; sys_platform != 'darwin' or platform_machine != 'x86_64'",
]

[project.scripts]
omnivoice-infer = "omnivoice.cli.infer:main"
omnivoice-infer-batch = "omnivoice.cli.infer_batch:main"
omnivoice-demo = "omnivoice.cli.demo:main"
omnivoice-dub = "omnivoice.cli.dub:main"

[project.urls]
Homepage = "https://github.com/debpalash/VoiceStudio"
Repository = "https://github.com/debpalash/VoiceStudio"
Documentation = "https://github.com/debpalash/VoiceStudio/tree/main/docs"
"Bug Tracker" = "https://github.com/debpalash/VoiceStudio/issues"
"Upstream TTS Model" = "https://github.com/k2-fsa/OmniVoice"

[tool.uv.sources]
# Install PyTorch with CUDA support on Linux/Windows (CUDA doesn't exist for Mac).
# NOTE: We must explicitly request them as `dependencies` above. These improved
# versions will not be selected if they're only third-party dependencies.
torch = [
  { index = "pytorch-cuda", marker = "platform_machine != 'aarch64' and platform_machine != 'arm64' and (sys_platform == 'linux' or sys_platform == 'win32')" },
]
torchaudio = [
  { index = "pytorch-cuda", marker = "platform_machine != 'aarch64' and platform_machine != 'arm64' and (sys_platform == 'linux' or sys_platform == 'win32')" },
]
torchvision = [
  { index = "pytorch-cuda", marker = "platform_machine != 'aarch64' and platform_machine != 'arm64' and (sys_platform == 'linux' or sys_platform == 'win32')" },
]

[[tool.uv.index]]
name = "pytorch-cuda"
# Use PyTorch built for NVIDIA Toolkit version 12.8.
# Available versions: https://pytorch.org/get-started/locally/
url = "https://download.pytorch.org/whl/cu128"
# Only use this index when explicitly requested by `tool.uv.sources`.
explicit = true

[tool.uv]
constraint-dependencies = [
    "mako>=1.3.12",
    "msgpack>=1.2.1",
    "pillow>=12.3.0",
    "pydantic-settings>=2.14.2",
    "pygments>=2.20.0",
    "starlette>=1.3.1",
    "torch==2.8.0",
    "torchaudio==2.8.0",
    "torchvision==0.23.0",
]

[tool.hatch.metadata]
# Needed so the KittenTTS wheel-URL dep in `project.dependencies` is accepted
# by hatchling's metadata validator. KittenTTS isn't on PyPI (dev preview),
# so pulling it via GH Releases URL is the only option today.
allow-direct-references = true

[tool.hatch.build.targets.sdist]
include = ["omnivoice"]

[tool.hatch.build.targets.wheel]
packages = ["omnivoice"]

[dependency-groups]
dev = [
    "httpx>=0.28.1",
    "pytest>=9.0.3",
    "pytest-asyncio>=1.3.0",
    "pytest-cov>=6.0",
    # Regenerates backend/worker/protocol/gen/ from worker_v1.proto. Dev-only:
    # the generated stubs are committed, so neither the installer nor Docker
    # needs protoc. tests/test_worker_protocol_gen.py fails if the two drift.
    "grpcio-tools>=1.60",
]

[tool.pytest.ini_options]
# Bare `pytest` would otherwise walk into `research/` (1.2 GB of vendored
# upstream projects, each with its own test_*.py that calls sys.exit at
# module level) and INTERNALERROR. `backend/tests/` still runs as its own CI
# session (see ci.yml) but no longer stubs sys.modules — its conftest.py sets
# a hermetic OMNIVOICE_DATA_DIR instead, so mixed invocations are safe too.
testpaths = ["tests"]
norecursedirs = [
    "research",
    "omnivoice/training",
    "omnivoice/eval",
    "frontend",
    "deploy",
    ".venv",
    "node_modules",
    "omnivoice_data",
    "backend/omnivoice_data",
]