Minimum sustained speech before an interim-triggered barge-in aborts the
agent's reply (pipeline mode) — measured from the utterance's first partial,
LiveKit's min_interruption_duration analog. A companion to
DEFAULT_MIN_BARGE_IN_WORDS: that one asks "is this enough words to be
an interruption", this one asks "has it lasted long enough to be speech at
all". Committed turns (STT finals) are never gated, so nothing the caller
actually said is lost — a gated barge-in only means the agent finishes its
sentence first.
Non-zero by default because the alternative is worse than the latency: room
noise and the tail of the agent's own audio both produce short interim
transcripts, and every one of them used to abandon a reply mid-word. Callers
heard the agent give up on its own sentences.
Minimum sustained speech before an interim-triggered barge-in aborts the agent's reply (pipeline mode) — measured from the utterance's first partial, LiveKit's
min_interruption_durationanalog. A companion to DEFAULT_MIN_BARGE_IN_WORDS: that one asks "is this enough words to be an interruption", this one asks "has it lasted long enough to be speech at all". Committed turns (STT finals) are never gated, so nothing the caller actually said is lost — a gated barge-in only means the agent finishes its sentence first.Non-zero by default because the alternative is worse than the latency: room noise and the tail of the agent's own audio both produce short interim transcripts, and every one of them used to abandon a reply mid-word. Callers heard the agent give up on its own sentences.