Bug report: Unreliable detection of speech offset

[gaoyingming]

This interface works fine for the onset of valid speech, but it has a delay about 100 ms for the offset of speech.

For the same frame, the function 'vad.is_speech' returns different decisions when it is repeatedly called. Sometimes 'True', sometimes 'False'. I found this phenomenon usually happened to the frames around the offset position.

[OmarHory]

Same is happening to me, aggressiveness is 3 and is_speech detects my noisy laptop mic as Speech, then sometimes it goes False.. My room is completely silent.
As compared when I use a professional Microphone it does really well, I am not sure what's the problem!

[PingAnPH]

I meet the same problem too. If I instantiate the vad once vad = webrtc.vad(), and repeatedly detect the same segment using vad.is_speech(), I will get different results. But when i instantiate the vad again,it will keep the same result at first invoked.
It seems that the parameter of vad is always changing, does anyone konw why?

[sparshsinha123]

@gaoyingming @OmarHory @PingAnPH @wiseman do we have any reason for this? Even I observed this in some experiments