PROTOCOL NOTES / 12

09 · Silence, speaking, feedback and voice-activated transmission

Participants: Audio, Engine, Peer

Implemented software · Physical qualification pending

Starting state

Active unmuted source; voice-activated transmission on by default (a persisted opt-out selects continuous transmission).

Ordered events

  1. Audio → Audio

    Analyze each processed 10-ms block: gated voice probability, encoder-input level, echo-like flag

  2. Audio → Engine

    Encode every 20-ms frame with Opus DTX off; hold paused frames as pre-roll

  3. Engine → Peer

    Signed source_status on transition and heartbeat repetition

  4. Peer → Engine

    Packet-index feedback reports delivery gaps separately from local lateness; two new windows of transport pressure reduce to 16 kbit/s; corroborated delivery gaps only cap at 24

  5. Audio → Peer

    Speech, onset or uncertainty resumes; held frames go first, the first with the RTP marker

Fields and messages

arrivalJitterMs and playoutTargetMs are distinct; voiceActivation default true; transmission state/revision/effectiveTimestamp/nextIndex; RTP time advances while sequence does not during suppression.

Authorization and invariants

Hints tied to admitted source/current SSRC never reset replay or suppress actual received media. Speaking uses qualified shared VAD hysteresis, independent of suppression and mixer gain; remote indications require authenticated fresh hints and recent decoded media.

Timers

2-s start-up; 1-s hangover; 20-ms analysis freshness; 30-ms callback gap; 5-frame (100-ms) pre-roll; hints fresh 3 s. Activity on at >=0.8 for 30 ms, off below 0.4 for 200 ms; activity-only publication at most 5/s.

Resource lifetime

Opt-out persists; VAD/codec policy lives in local audio session; keys and extended packet indexes survive silence.

What the rider sees

Settings switch "Transmit only while talking" (on by default); activity remains independent of bandwidth savings.

Failure branches

Unknown/stale analysis transmits; paused frames send nothing; timestamp-unrelated FEC is skipped; bounded PLC ends in silence.

Resulting state

Continuous mode sends every encoded frame; voice-activated mode does not send paused frames and, on resume, sends up to five held frames first.

Evidence boundary

Continuous/voice two-mode annotated corpus before encoding and through production DSP/Opus, spoken-digit onset test, policy/sender/receiver/activity tests and multi-hop loss/jitter regression. No field intelligibility claim.

← All scenarios

Authority: RWCP v2 reviewed scenario catalogue, scenario 09. Public examples are synthetic. This guide is maintained against the private implementation; it does not imply access to its source.