Palash Debnath
ca8a2e8eb8
The audiobook importer accepted .txt/.md/.epub but not PDF — the single most common "ebook in" format. Add a pure `pdf_to_chapter_script(data)` that extracts the text layer page-by-page and runs it through the existing chapterizer, so PDFs land in the same `# Heading` + body grammar EPUB and plaintext already produce (one front door onto the unchanged render pipeline). - Dep: `pypdf>=4.0` — pure-Python, MIT, zero native deps, so PDF import behaves identically on macOS/Windows/Linux (default-feature cross-platform rule). EPUB + plaintext stay stdlib-only; only PDF needs a real parser. - Robustness, surfaced as actionable 400s rather than silent empty imports: corrupt file, password-protected (empty-password decrypt attempted first), scanned/image-only (no text layer → clear "scanned PDF" message), and a page-count ceiling. A single unparseable page is skipped, not fatal. - Route: `.pdf` branch in audiobook_import; frontend accept filter + api-client doc updated to `.txt,.md,.epub,.pdf`. tests/test_longform_import.py: 5 PDF cases (extract+chapterize, no-marker single chapter, corrupt, image-only, page-cap) using a hand-built in-memory PDF — no PDF-authoring test dep, mirroring the in-memory-EPUB approach. 16 passed; frontend suite 401; CJK guard green.
238 lines
9.6 KiB
TOML
238 lines
9.6 KiB
TOML
[build-system]
|
||
requires = ["hatchling"]
|
||
build-backend = "hatchling.build"
|
||
|
||
[project]
|
||
name = "omnivoice"
|
||
version = "0.3.6"
|
||
description = "OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models"
|
||
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 OmniVoice@palash.dev. 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 = "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",
|
||
"transformers>=5.3.0",
|
||
"accelerate",
|
||
"pydub",
|
||
"gradio",
|
||
"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",
|
||
"pyinstaller>=6.19.0",
|
||
"imageio-ffmpeg>=0.6.0",
|
||
"pedalboard>=0.9.14",
|
||
# 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'",
|
||
"demucs>=4.0.1",
|
||
"yt-dlp>=2024.12.13",
|
||
"alembic>=1.13",
|
||
# 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.
|
||
"fastapi",
|
||
"scalar-fastapi",
|
||
"uvicorn",
|
||
"python-multipart",
|
||
# 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>=1.2",
|
||
# 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 — see .planning/quick/260613-fdl-fast-model-downloads/SPIKE).
|
||
# 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>=4.0",
|
||
]
|
||
|
||
[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",
|
||
"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",
|
||
]
|
||
|
||
[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/k2-fsa/OmniVoice"
|
||
Repository = "https://github.com/k2-fsa/OmniVoice"
|
||
"Bug Tracker" = "https://github.com/k2-fsa/OmniVoice/issues"
|
||
|
||
[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 = "sys_platform == 'linux' or sys_platform == 'win32'" },
|
||
]
|
||
torchaudio = [
|
||
{ index = "pytorch-cuda", marker = "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 = [
|
||
"torch==2.8.0",
|
||
"torchaudio==2.8.0",
|
||
]
|
||
|
||
[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",
|
||
]
|
||
|
||
[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/` runs separately because
|
||
# it stubs `core.config` in sys.modules, which pollutes import state for
|
||
# other modules in the same session.
|
||
testpaths = ["tests"]
|
||
norecursedirs = [
|
||
"research",
|
||
"omnivoice/training",
|
||
"omnivoice/eval",
|
||
"frontend",
|
||
"deploy",
|
||
".venv",
|
||
"node_modules",
|
||
"omnivoice_data",
|
||
"backend/omnivoice_data",
|
||
]
|