The Real-time Speech-to-Text documentation says “The Voice Activity Detection (VAD) endpoint determines whether audio contains speech or silence” - well, it’s deciding my AudioHook stream is silent when it absolutely isn’t, roughly 1 in 10 calls. We’re using the WebSocket connection, obviously, and the audio.vad.enabled parameter is true on the AudioHook activity.
It’s like the signal gets dropped for half a second, then picks back up, which ruins the agent assist summary; we spun up a quick Lambda to re-request the stream on disconnect, but that’s just a band-aid. I’m seeing this consistently between 9am and 11am EST, which makes me think regional latency is still the root cause, honestly.
Here’s a sample log snippet during a call where VAD erroneously reported silence:
{
"event": "audio.vad.detected",
"timestamp": "2024-02-29T10:32:15.123Z",
"vad_result": "SILENCE",
"confidence": 0.95
}
The official docs also mention “Ensure that the audio source is consistently transmitting audio data,” so I’m checking that now, but it feels like the problem isn’t our audio source.