vadonnx

Load arbitrary Voice Activity Detection models behind a single, unified API. Every model runs through ONNX Runtime. You get one streaming and batch interface, one audio-format story, and pluggable models.

from vadonnx import load_vad

vad = load_vad("silero")                              # bundled, works fully offline
prob = vad.process_chunk(pcm_bytes)                   # streaming → float in [0, 1]
segments = vad.get_speech_segments(audio, sample_rate=16000)
# -> [SpeechSegment(start=0.32, end=2.27), SpeechSegment(start=3.27, end=4.45), ...]

Why

Every VAD ships its own loader, audio format, feature pipeline, and state handling. vadonnx hides that behind one VADModel interface. Feed it audio (raw int16 bytes, numpy arrays, any sample rate) and get back per-frame speech probabilities or ready-made speech segments. An IOSignature describes each model declaratively, so a single generic engine drives most of them, and you can point the same API at any custom .onnx file.

• Lightweight runtime: only numpy, onnxruntime, and huggingface_hub.
• Offline by default: a small Silero model is bundled in the wheel.
• Streaming and batch: process_chunk() for live audio, get_speech_segments() and probabilities() for whole buffers.
• Bring your own model: load any ONNX VAD by path or URL with a signature.
• Extensible: third parties register backends and models through entry points.

Install

uv pip install vadonnx          # runtime (numpy + onnxruntime + huggingface_hub)
uv pip install "vadonnx[mic]"   # + microphone examples

Models

See docs/backends.md for per-model detail and the benchmark for measured comparisons across datasets, including WebRTC and energy baselines.

Models other than the bundled Silero are downloaded on first use from the TigreGotico HuggingFace org and cached under $XDG_DATA_HOME/vadonnx. The PulseVAD models download from a pinned commit of the upstream GitHub repository, and a sha256 digest checks each file. See docs/backends.md for per-model detail and parity notes.

CLI

vadonnx list                       # list available models
vadonnx probe silero               # print a model's ONNX input/output signature
vadonnx segment speech.wav         # print detected speech segments of a WAV

Documentation

• Quickstart
• Streaming
• Custom models & IOSignature
• Backends & parity notes
• Plugins
• Model conversion
• Licensing
• API reference

Related projects

• TigreGotico/phoonnx: multilingual phonemization and ONNX text-to-speech, the TTS counterpart to vadonnx.
• TigreGotico/onnx-asr: offline speech recognition on ONNX models.

License

Apache-2.0. Bundled and downloaded model weights retain their upstream licenses. See docs/licensing.md.