Thanks — I’m responding to your question about preventing filler words (e.g., “hmm”, “ok”, “aha”, “umm”) from interrupting the bot. Yes — LiveKit provides adaptive interruption handling that’s designed to ignore brief backchannels and filler words so the agent doesn’t yield its turn unnecessarily. What you can do now • Enable adaptive interruption handling (requires Python SDK v1.5.0+ or Node SDK v1.2.0+, VAD on, and an STT model with aligned transcripts). • Adaptive mode uses audio+prosody to distinguish real barge-ins from backchanneling and is enabled by default for LiveKit Cloud agents and dev mode when the prerequisites are met. • If you prefer simpler behavior, use VAD-only mode or tune the interruption thresholds: ◦

min_duration

— minimum speech duration (seconds) to count as an interruption (default 0.5s). ◦

min_words

— minimum transcribed words required to trigger an interruption (default 0). Quick examples 1. Python

session = AgentSession(
    turn_handling=TurnHandlingOptions(
        interruption={"mode": "adaptive"},
    ),
)

1. TypeScript

const session = new voice.AgentSession({
  turnHandling: {
    interruption: { mode: 'adaptive' },
  },
});

If you need filler words to be even less likely to interrupt, increase

min_duration

(e.g., 0.6–0.8s) and/or set

min_words

to 1 so very short/noisy utterances aren’t treated as interruptions. If you want, tell me: • Which SDK (Python or Node) you’re using, • Whether you’re using LiveKit Cloud or a local agent, and I’ll suggest concrete settings (values) to tune for your use case. Sources: Adaptive interruption handling | LiveKit Documentation | Solving unwanted interruptions with Adaptive Interruption Handling | LiveKi... | Turns overview | LiveKit Documentation, with interactive elements