Detect Voice - Suppress Non-Speech not working

[lachimsun]

It seems as if it tries to constantly record random sounds and then tries to translate them (using whisper). It makes no difference if I speak or make some other noise.

[LostRuins]

Non speech tokens should be suppressed though, you won't see them in your output

The sound detection part is separate. That one is done in browser.

[lachimsun]

Unfortunately it generates some random words for me. On the other hand if I use e.g. Open-LLM-VTuber speech recognition works perfectly (using firefox).

[LostRuins]

Hmm strange. Can you try load a whisper model, upload a mp3 voice clip to the Lite UI, and then trigger the transcribe in the UI, does it get the speech text right?

[lachimsun]

Transcription of continuous voice clips works well, but I think the problem occurs when we stop speaking for a while, as if the start/end of a speaking isn't detected.

[LostRuins]

Okay in that case, instead of Detect Voice, try using Push To Talk or Toggle To Talk

[lachimsun]

Thanks for your replies, although I was hoping hands-free operation would be possible.

[LostRuins]

It is, I don't know why the VAD ( Voice activated detection) is not working for your browser/mic. perhaps the background volume is too inconsistent.