Voice Activity Detection

VAD monitors an endpoint's incoming audio and emits events when speech starts or silence persists. It is completely independent of recording — they can be used separately or together.

Use Cases

• IVR / Voicemail: Detect when a caller finishes speaking after a prompt
• LLM Streaming: Know when the user is done talking before sending audio to an LLM
• Call Quality: Monitor whether participants are actively speaking

vad.start

{
  "id": "1",
  "method": "vad.start",
  "params": {
    "endpoint_id": "...",
    "silence_interval_ms": 1000,
    "speech_threshold": 0.5
  }
}

Param	Type	Default	Description
endpoint_id	string	required	Endpoint to monitor
silence_interval_ms	u32	1000	How often to emit silence events (ms, 100-3600000)
speech_threshold	f32	0.5	earshot score threshold (0.0-1.0)

vad.stop

{"id":"2","method":"vad.stop","params":{"endpoint_id":"..."}}

Events

vad.speech_started

Emitted when speech is detected after a period of silence.

{"event":"vad.speech_started","data":{"endpoint_id":"..."}}

vad.silence

Emitted periodically while silence persists (every silence_interval_ms). The first emission occurs after silence_interval_ms of continuous silence.

Speech-to-silence transitions retain a 320 ms hangover. Silence duration and event cadence use the greater of decoded audio time and monotonic elapsed time since the last detected speech/transition. Quiet packets do not reset that elapsed-time clock. This preserves audio-time behavior for buffered input while allowing Opus DTX (for example, 20 ms packets arriving every 160 ms) and complete packet cessation to reach the configured interval promptly. A timeout transition never emits an immediate zero-duration silence event; the full interval must elapse after the hangover.

{"event":"vad.silence","data":{"endpoint_id":"...","silence_duration_ms":3000}}

vad.error

Emitted when the VAD decoder cannot be created for the endpoint's codec.

{"event":"vad.error","data":{"endpoint_id":"...","error":"VAD decoder creation failed: ..."}}

IVR Pattern

For an IVR scenario where a greeting plays before listening for a response:

1. Start file playback (greeting prompt)
2. Start VAD on the caller's endpoint
3. Ignore vad.silence events while the greeting is still playing
4. When endpoint.file.finished fires, start paying attention to VAD events
5. Wait for vad.speech_started (caller begins responding)
6. Wait for vad.silence with sufficient silence_duration_ms (caller finished)
7. Process the response

This pattern keeps the VAD logic simple in rtpbridge while giving the control application full flexibility.