Silero VAD v6.2.0: GGUF for NeMo-Speech.cpp

Silero VAD v6.2.0, a small voice activity detector by the Silero Team, converted to the GGUF that NeMo-Speech.cpp v0.3.0 loads as its VAD companion. NeMo-Speech.cpp uses it to mask silence before the speech recogniser sees it, and to end utterances.

It was converted for Goscriber, an offline push-to-talk dictation app, and works with any NeMo-Speech.cpp runtime. This is not a Silero Team or NVIDIA release.

Files

Use

Pass it next to the ASR model and turn masking on:

nemo-speech transcribe audio.wav --model parakeet-tdt-0.6b-v2.q8_0.gguf \
  --vad-model silero-vad-v6.2.0.gguf --vad-masking

Through the C ABI, set model_path and enable_masking in nemo_speech_asr_vad_config. The file holds the 16 kHz model only, and it is not the whisper.cpp Silero GGML format.

How it was made

• Source: the silero-vad 6.2.0 package from PyPI (wheel SHA-256 8a31f1e58e73cdc20820d842f757e678e0ae1b0ee1cb7b8aa50f20f8fd7f7638), which carries the model weights.

• Converter: convert_model.py from NeMo-Speech.cpp v0.3.0 (commit 20999cc9), on Python 3.12 with hash-pinned packages.

• Command:

  python convert_model.py silero --silero-version 6.2.0 --outfile silero-vad-v6.2.0.gguf
  

• Reproduce: scripts/convert-models.sh in Goscriber builds this file and Parakeet TDT 0.6B v2 from scratch in an isolated environment, then checks their SHA-256 against these. See docs/models/gguf.md.

The weights are copied as they are, in F32: nothing is quantised or retrained.

License and attribution

Silero VAD is © 2020-present Silero Team and released under the MIT License. This GGUF is the same model in another file format and keeps that licence: see LICENSE, copied from the silero-vad 6.2.0 package.