CoolFace
Modelpublic

pat229988/silero-vad-16k-tflite

sourceHugging Facemitupdated 3mo agoView on Hugging Face
0likes26downloads
Model Card

Silero VAD 16 kHz TFLite

This repository contains a 16 kHz Silero VAD model converted to TFLite/LiteRT for local streaming voice activity detection.

Files

FilePurpose
silero_vad_16k.tfliteRuntime TFLite flatbuffer.
silero_vad_16k.tflite.manifest.jsonConversion/runtime manifest.
artifact_manifest.jsonClean upload metadata, checksum, and source lineage.

Source lineage

The original Silero VAD weights were sourced from the upstream GitHub data folder provided here:

<https://github.com/snakers4/silero-vad/tree/master/src/silero_vad/data>

The local source artifact referenced during conversion was:

text
external/silero-vad/src/silero_vad/data/silero_vad_16k.safetensors

Because the source is a GitHub-hosted artifact rather than a Hugging Face model repo, it is not listed in the Hugging Face base_model metadata field. The upstream source link is included here and in artifact_manifest.json.

Runtime contract

  • —Sample rate: 16000 Hz
  • —Streaming input: 64 samples of left context + current 512 sample chunk (576 total samples)
  • —Recurrent state input/output shape: [2, 1, 128]
  • —Probability output shape: [1, 1]
  • —Common decision threshold: 0.5 (adjust per environment and application)

Limitations

  • —This is a converted artifact.
  • —VAD probability thresholds are environment-dependent. Validate against your microphones and background-noise conditions.