v6.2

a16z-infra/llm-app-stackv6.2Nov 6, 2025by snakers4

AI Summary

This release focuses on enhancing the Voice Activity Detection (VAD) model's robustness by reworking the training paradigm and improving stability on out-of-distribution data.

Key Highlights

Reworked VAD training paradigm
Improved stability on OOD / rare / strange / unique data
Significant quality improvements on edge cases (unusual, child, cartoon voices, muted speech)
Fixed type hint for min_silence_at_max_speech

New Features

Enhanced handling of lower quality phone calls
Improved stability on various known edge cases

Full Release Notes

- VAD training paradigm reworked;
- Overall slight quality improvement (no metrics update);
- Higher stability on OOD / rare / strange / unique data;
- Significant quality improvements on various known edge cases:
  - Unusual voices
  - Child voices
  - Cartoon voices
  - Muted voices
  - Muted speech
  - Lower quality phone calls

## What's Changed

* Fix type hint for min_silence_at_max_speech (float -> int) by @Purfview in https://github.com/snakers4/silero-vad/pull/714
* Adamnsandle by @adamnsandle in https://github.com/snakers4/silero-vad/pull/717
* Adamnsandle by @adamnsandle in https://github.com/snakers4/silero-vad/pull/719

**Full Changelog**: https://github.com/snakers4/silero-vad/compare/v6.1...v6.2