Higher confidence scores for turn_completion only after 1200ms of silence at end of audio sample.

[atharvabagde]

I have deployed the smart-turn onnx model in rust and have been using it for detecting "SpeechEND" in my VAD service. By default the VAD service I am using (TEN VAD) has a fallback for SpeechEND after 800ms of silence. For all the input audio samples, the fallback is triggered while the smart_turn probability score is still at around 0.4 for end_of_turn.

Wanted to know if there is something wrong in my implementation or is there an average silence of around 1-1.2s at the end for all end_of_turn annotated samples.

PS:: Is there any drawback for using hard-coded 800ms of SpeechEND threshold. Please let me know of any suggestions/feedback you have on that

[marcus-daily]

Hi @atharvabagde, I'm not sure I understand the question. Smart Turn shouldn't be affected by the amount of silence at the end of the sample. Are you seeing different results with different VAD parameters?

Note that the full context of the current turn (up to 8 seconds of audio) should be provided to Smart Turn, in your "SpeechEND" callback from VAD. Smart Turn is not a streaming model but instead works on large chunks of audio.

[atharvabagde]

Yess...I am sending full content (8s) to the model.
My question is that for the training audio samples that are annotated as "Complete", do they usually have 1-1.2s of silence at the end. Because when I send a 8s audio sample with upto 800ms of silence at the end....smart_turn gives very low probability.

Eg: Speech...800ms of silence --> 0.4 probability score
Speech....1.3s of silence --> 0.89 probability score