Interface AMDOptions

Hard ceiling for the entire detection. After this, settle with whatever evidence exists.

Silence after a short greeting before settling as HUMAN.

Speech longer than this is treated as machine-like (skips the short-greeting heuristic).

LLM used to classify call greetings.

• LLM instance: used as-is (caller-owned; AMD will not close it).
• string: treated as a Cloud Inference model id (e.g. 'openai/gpt-4o-mini') and an inference LLM is constructed (AMD-owned).
• undefined (default): auto-select — if LiveKit Cloud inference credentials are available in the environment, uses 'google/gemini-3.1-flash-lite' via the inference gateway; otherwise falls back to the session's own LLM.

Silence after machine-like speech before opening the silence gate.

Fallback end-of-turn delay (ms). When the session turn detector never commits a turn, this synthetic backstop, armed when speech ends, sets the end-of-turn so a gated verdict can still emit. Defaults to the running session activity's endpointing maxDelay (so the backstop tracks the real turn detector), or DEFAULT_MAX_ENDPOINTING_DELAY_MS when no activity is available. Mirrors python max_endpointing_delay.

If no speech is heard within this window, settle as UNCERTAIN (not a machine, so no interrupt).

Restrict AMD to a specific participant. Used to filter the waitForTrackPublication gate (see python detector.py) and span attribution. When unset, AMD binds to whichever participant the session is linked to.

Override the AMD classification system prompt.

Dedicated STT used to transcribe call audio for AMD.

• STT instance: used as-is (caller-owned; AMD will not close it).
• string: treated as a Cloud Inference model id (e.g. 'cartesia/ink-whisper') and an inference STT is constructed (AMD-owned).
• undefined (default): auto-select — if LiveKit Cloud inference credentials are available in the environment, uses 'cartesia/ink-whisper' via the inference gateway; otherwise reuses the session's existing STT transcripts.

If true, do not log a warning when the resolved LLM is not among the bundled AMD-tested model strings. Has no effect on classification behavior.

If true, once any speech has been heard the detectionTimeout no longer forces emission — AMD keeps waiting for post-speech silence and a positive end-of-turn from the session's turn detector before emitting. Useful for outbound voicemail flows where leaving a message early would overlap the greeting. noSpeechTimeout (uncertain) still fires normally (no audio at all means there is nothing to wait for). Defaults to false. Mirrors python detector.py wait_until_finished.