bitsydarel/silero-vad-onnx

Silero VAD — ONNX mirror

This repository mirrors the official Silero VAD ONNX exports under versioned filenames so multiple releases can be hosted side-by-side and consumed by SDKs that pin to a specific revision.

Files

File	Upstream version	Source	SHA-256
silero_vad_v6.2.1.onnx	v6.2.1 (released 2025-02-24)	snakers4/silero-vad@v6.2.1	1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3

Model spec (v6.2.1)

• Sample rate: 16 kHz (fixed)
• Chunk size: 512 samples (32 ms @ 16 kHz)
• Inputs:
  • input: float32 [1, 512] audio chunk
  • state: float32 [2, 1, 128] LSTM hidden + cell state
  • sr: int64 scalar (16000)
• Outputs:
  • output: float32 [1, 1] speech probability
  • stateN: float32 [2, 1, 128] updated LSTM state

The LSTM state must be initialised to zeros for the first chunk of every stream and threaded through subsequent chunks. Reset between independent streams to avoid leaking hidden state across turns.

License

MIT. See LICENSE — verbatim copy of the upstream snakers4/silero-vad@v6.2.1/LICENSE. © 2020-present Silero Team.

Why a mirror

The Silero repository hosts only the latest model under a single filename (silero_vad.onnx). Consumers that need to pin to a specific version for reproducibility (SHA-verified downloads, multi-version SDK rollouts) need versioned filenames. This mirror provides that without modifying the artifacts — the SHA-256 above matches the upstream file byte-for-byte.

Consumers

• BDAIAssistantSDK — pins silero_vad_v6.2.1.onnx via SileroOnnxVADConfiguration.