vllm_omni.engine.duplex.session.overlap_policy ¶

How one duplex session reacts to input that arrives while it is speaking.

Short acknowledgement, barge-in, or keep listening: the rules read the session config and the incoming append, and nothing else. Extracted from DuplexSessionRunner because they were a decision function wearing a method's clothes -- the only runner state they ever touched was session -- and because reaching them previously meant driving a whole append through the runner.

decide ¶

decide(
    session: DuplexEngineSession,
    event: dict[str, object],
    payload: dict[str, object],
    *,
    auto_responds: bool,
) -> dict[str, object]

defer_unsupported_barge_in ¶

defer_unsupported_barge_in(
    session: DuplexEngineSession,
    *,
    duration_ms: int,
    is_speech: bool,
) -> dict[str, object]

event_requests_barge_in ¶

event_requests_barge_in(
    event: Mapping[str, object],
) -> bool

input_audio_duration_ms ¶

input_audio_duration_ms(
    event: dict[str, object], payload: dict[str, object]
) -> int

input_looks_like_speech ¶

input_looks_like_speech(
    session: DuplexEngineSession,
    event: dict[str, object],
    payload: dict[str, object],
) -> bool

is_short_ack_transcript_hint ¶

is_short_ack_transcript_hint(
    event: dict[str, object], payload: dict[str, object]
) -> bool

merge_audio_payloads ¶

merge_audio_payloads(
    first: dict[str, object], second: dict[str, object]
) -> dict[str, object]

should_force_listen_for_auto_response_overlap ¶

should_force_listen_for_auto_response_overlap(
    event: dict[str, object],
    payload: dict[str, object],
    *,
    auto_responds: bool,
) -> bool

should_force_listen_for_short_commit ¶

should_force_listen_for_short_commit(
    session: DuplexEngineSession,
    event: dict[str, object],
    payload: dict[str, object],
) -> bool

vad_speech_started ¶

vad_speech_started(
    event: Mapping[str, object],
    payload: Mapping[str, object],
) -> bool | None