❓ Questions / Help / Support - Silero VAD misses silence boundaries detected by FireRedVAD in Russian speech

[chuducnam]

First of all, thank you for creating and maintaining Silero VAD and making it freely available to the community.

I tested Silero VAD on three Russian interview recordings, totaling about 459 seconds, and compared the results with FireRedVAD.

FireRedVAD detected 18 silence boundaries inside regions that Silero returned as continuous speech. At those boundaries, FireRedVAD remained below its threshold for 310–600 ms, exceeding the configured minimum silence duration of 250 ms.

Over the same intervals, Silero remained below its default exit threshold of 0.35 for only 0–224 ms, and in several cases did not fall below 0.35 at all.

As a result, Silero sometimes merged multiple speech turns into continuous segments lasting 64–71 seconds despite audible pauses longer than 250 ms.

This concerns silence-boundary detection, not speaker diarization.

Could you please clarify whether this is expected behavior, a known model limitation, or a possible issue that should be investigated?

I have attached the WAV files, JSON reports, and configurations needed for analysis.

Thank you for your time and for your work on this project.

audio.zip

vad_reports.zip

[adamnsandle]

Thanks a lot for the detailed report!

What you're seeing is actually a known limitation of the current Silero VAD. In dense, continuous speech (such as interviews or conversations with short pauses), the model sometimes fails to separate adjacent speech segments and ends up merging them into a single long segment. Improving speech boundary detection in these scenarios is something we plan to address in future releases.

From our experience, FireRedVAD indeed tends to detect short pauses inside dense speech better than Silero. However, the trade-off is that it is noticeably less robust to background noise and other acoustic artifacts. Silero was intentionally tuned to be more conservative, which unfortunately can result in missed short silence intervals like the ones you observed.

For the current model, the main thing you can experiment with is the threshold and especially the neg_threshold parameters. On clean audio, increasing neg_threshold (even up to around 0.8) often makes the model much more willing to close speech segments on short pauses. The optimal values are highly application-dependent, but this is probably the first thing we'd recommend trying.

[chuducnam]

Thank you for the SUPER-FAST and detailed response!

I really appreciate the explanation and the recommendation regarding neg_threshold.

I hope to see an improved version of Silero VAD in the near future.

Have a great day!