You can deploy this model and start earning money today!

idle-intelligence/
silero-vad-v5-safetensors

Name: silero-vad-v5-safetensors
Availability: InStock

silero-vad

voice-activity-detection

Views

No views yet

Silero VAD v5 (safetensors)

Silero Voice Activity Detection v5, converted from ONNX to safetensors format for use with candle / Rust inference.

Original model: snakers4/silero-vad (MIT License)

Model Details


Task	Voice Activity Detection
Input	16 kHz mono PCM, 512 samples per call (32 ms)
Output	Speech probability ∈ [0, 1]
Size	1.2 MB
Format	safetensors (f32)

Architecture

Input: [batch, 576] (64 context + 512 audio samples)
  → Reflection pad (64 samples, right only) → [batch, 1, 640]
  → Conv1d(1, 258, k=256, stride=128) → STFT → magnitude [batch, 129, 4]
  → Conv1d(129, 128, k=3, pad=1) + ReLU
  → Conv1d(128, 64, k=3, stride=2, pad=1) + ReLU
  → Conv1d(64, 64, k=3, stride=2, pad=1) + ReLU
  → Conv1d(64, 128, k=3, pad=1) + ReLU
  → LSTM(128, 128) → ReLU
  → Conv1d(128, 1, k=1) + Sigmoid → probability

The model is stateful — LSTM hidden/cell state and a 64-sample context buffer carry across calls.

Tensor Names

Tensor	Shape
`stft.conv.weight`	[258, 1, 256]
`encoder.{0-3}.conv.weight`	Conv weights (BN-fused / reparameterized)
`encoder.{0-3}.conv.bias`	Conv biases
`decoder.lstm.weight_ih_l0`	[512, 128]
`decoder.lstm.weight_hh_l0`	[512, 128]
`decoder.lstm.bias_ih_l0`	[512]
`decoder.lstm.bias_hh_l0`	[512]
`decoder.output.weight`	[1, 128, 1]
`decoder.output.bias`	[1]

Usage

Used by vad-rs for client-side speech detection in the browser (via WASM + candle-core). See the original silero-vad repository for Python/ONNX usage examples and documentation.

Thresholds

Parameter	Default
Speech start	probability ≥ 0.5
Speech end	probability < 0.35

License

MIT (same as original Silero VAD)

idle-intelligence/silero-vad-v5-safetensors