The VAD receives 8kHz, 16-bit, mono, signed-linear PCM data over a TCP socket, running it through the Silero VAD ML model, and returns for each processed block a one-byte unsigned integer (00-FF) for the confidence that voice is present.
Generally, we seem to be able to use fairly high confidences as a threshold:
>97% (> 0xF7).
Usage note: the model is intended as an aggregate, so whenever the client side determines that it has sufficient data to indicate a voice, it should terminate the connection, dialing again when it wants to detect voice again, even if that is during the same telephone conversation.
Samples may be tested with ffmpeg and nc, like so:
$ ffmpeg -hide_banner -loglevel error -re -i test.wav -f s16le - | nc localhost 3030 |hexdump