How to properly handle the probability produced by ten-vad?

[Josscii]

Hi, I've got the probability of some audio data, how to determine the voice part, is there a formal algorithm？or we need to implement one based on our own needs.

Thanks for open sourcing this!

[Ziyi6]

Our VAD returns voice probs frame-wise, meaning that for each frame, the model would return a prob indicating the speech activity. I think your case is: given a speech segment, the VAD model may return prob smaller than threshold in some frame and what should we do.

If this is the case, then you may need to do a smoothing either for the output probs or the indicator output, e.g. moving average or median filtering.

[bzikst]

You can adapt code from any other vad, for example: https://github.com/snakers4/silero-vad/blob/94811cbe1207ec24bc0f5370b895364b8934936f/src/silero_vad/utils_vad.py#L191

Or maybe write your own for a specific use case.

[Josscii]

Thanks, that's what I mean in the first place, as I see silero implemented some logic to determine what should be considered as speech based on minimal duration like that.