wait for user silence before speaking

[chenghao-mou]

Speech will wait for both authorization and user silence.

User silence is defined as:

• VAD: not speaking
• VAD: speaking but has raw_accumulated_silence

[theomonnom]

If the user is in a noisy environment, we may never trigger the speech? VAD is a very unstable signal

[longcw]

we may want a threshold of raw_accumulated_speech here?

[chenghao-mou]

Right, we definitely don't want it to be speaking between user words.

[longcw]

If the user is in a noisy environment, we may never trigger the speech? VAD is a very unstable signal

I think it's fine, if VAD has a lot of false positives in a noisy env, the agent speech will still be interrupted all the time even without this waiting mechanism. we can add an option to make the threshold configurable, maybe disabled by default right now.

[chenghao-mou]

If the user is in a noisy environment, we may never trigger the speech? VAD is a very unstable signal

I think it's fine, if VAD has a lot of false positives in a noisy env, the agent speech will still be interrupted all the time even without this waiting mechanism. we can add an option to make the threshold configurable, maybe disabled by default right now.

How about min_response_silence: how long the user silence the agent should wait for before it can respond, which should be shorter than min_endpointing_delay.