Model Card for Silero VAD (IB-Robot)

Silero VAD (voice activity detection) packaged for the IB-Robot framework as a standalone reusable bundle. It is a speech-domain base component consumed by several businesses: speech_direction DOA gray-region gating and ASR endpoint detection.

Deployments

deployment	backend	artifacts	notes
`ascend_310p`	Ascend ACL (Ascend310P1)	`silero_vad_v6_310p_mixed16.om`	sample rate folded to a constant
`torch_cpu`	ONNX Runtime CPU (torch profile)	— (ONNX in `assets/`)	official v6 ONNX

Shared fixed ABI (host.silero.*): audio [1,576] (512-sample chunk + 64 context @16 kHz) and LSTM state [2,1,128] in; prob [1,1] and state_out out. Stream execution contract (stateful, runtime_exclusive state bank, max_open_streams: 1). The ONNX sr input is fed the 16000 constant by the host runner, matching the OM's constant folding.

Weights provenance

assets/silero_vad.onnx = snakers4/silero-vad master v6 ONNX (sha256 1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3)
The 310P OM was converted from the same v6 model (fixed-ABI export)
Pinned in assets/adapter.json and the download script

Repository Structure

inference_manifest.json — deployment routing (schema v3, stream state contract)
assets/adapter.json — identity + checkpoint digest + frame constants
assets/silero_vad.onnx — official v6 ONNX (2.3 MB)
artifacts/ascend/ascend_310p/silero_vad_v6_310p_mixed16.om — 310P OM

Usage

Regenerate the bundle locally:

./scripts/download_speech_direction_models.sh --silero-only
python3 -m voice_asr_service.package_silero_vad_bundle \
  --bundle-root models/silero-vad --workspace .

Consumers resolve artifacts through the manifest (speech_direction silero_vad_inference_bundle parameter, default models/silero-vad); the unified inference runtime registers the tensor_model/silero_vad/vad identity for session construction.

License

Upstream snakers4/silero-vad is MIT-licensed; packaging follows the same license.

Downloads last month: -; Downloads are not tracked for this model. How to track