Experimental VAD/DTX

An experiment, disabled by default

The default media path sends continuous codec output. Experimental VAD/DTX can reduce traffic during qualified quiet periods, but its physical speech quality is unqualified. It is not marketed as a proven battery or wind-noise improvement.

Silence needs positive evidence

The VAD model consumes postprocessed FloatS16 audio in 10 ms blocks. Policy combines speech probability with acoustic measurements, startup protection, hangover, quiet qualification, and availability checks. Weak, stale, unknown, or unavailable evidence keeps transmission conservative.

Every active, unmuted frame still reaches the encoder. The sender suppresses only confirmed Opus no-transmit output and sends comfort updates as required. There is no intentional pre-roll or added packetization delay, and no pre-roll recovery if classification misses an onset.

A quiet source still has a clock

RTP timestamps, marker behavior, sequence numbers, rollover counters, and replay windows must remain coherent. Signed transmission hints help receivers distinguish intentional silence from packet loss. Source resynchronization can advance progress across a partition but must not rewind replay state or permit same-SSRC key replacement.

Resume without replaying a silence backlog

Receive policy uses source timing to choose FEC and bounded concealment, skips inappropriate recovery across intentional gaps, and avoids queued silence. Direct ping/pong liveness continues independently, so an intentionally quiet source is not considered a dead transport. Host corpus checks are separate from helmet, wind, and simultaneous-speaker qualification.