OptionalCoverageWhat a turn's input is allowed to include. Distinct from DETECTION (when a turn ends): coverage is WHAT rides in it.
Gemini 3.8 Live defaults to audioActivityAndAllVideo, which ships every video frame to the
model by default — billed and consuming context. Declaring coverage explicitly keeps that a
decision rather than an inherited default. Profiles without a mapping ignore it.
OptionalEagernessThe cross-driver patient (low) / normal (auto) / eager (high) turn-taking dial; ignored by drivers without a
mapping. OpenAI and Eleven Realtime: semantic-VAD eagerness, applied only with Mode: 'semanticVad'. xAI: an
end-of-turn silence ladder (off when any layer sets SilenceDurationMs). Gemini: end-of-speech sensitivity.
ElevenLabs Agents: turn eagerness.
OptionalModeThe normalized mode; absent = 'default'.
OptionalSilenceServer-VAD trailing-silence duration in ms; ignored by profiles without a mapping.
OptionalThresholdServer-VAD activation threshold (0–1); ignored by profiles without a mapping. A profile whose provider has coarser controls maps it as an approximate value: Gemini Live has two start-of-speech buckets, so below 0.5 is START_SENSITIVITY_HIGH and 0.5 or more is START_SENSITIVITY_LOW.
Normalized turn-detection settings. Every field optional; absent fields contribute nothing. Profiles translate to their native wire fields and ignore what they have no mapping for.