https://livekit.io logo
Join Slack
Powered by
```Adaptive Interruption Handling The headline fea...
# ask-ai
b
Copy code
Adaptive Interruption Handling
The headline feature of v1.5.0: an audio-based ML model that distinguishes genuine user interruptions from incidental sounds like backchannels ("mm-hmm"), coughs, sighs, or background noise. Enabled by default — no configuration needed.
this is from the release notes of agents version 1.5.0 so when detecting a false interrupt 1. while the agent is speaking state 2. while the agent is thinking state what exactly happens how does the agent behave in these cases
✅ 1
t
Thanks for reaching out, Mohammed. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.
Thanks — this is a great question about how the new adaptive interruption handling works in v1.5.0. I’ll address what happens for a false interrupt (an audio event that produces an empty transcription, e.g., cough or “mm‑hmm”) in both the agent’s speaking and thinking states, and what you can configure. Short answer • While the agent is speaking: the agent immediately stops on detected voice activity. If 
resume_false_interruption
is true, the agent will resume from where it left off after silence for `false_interruption_timeout`; if false, the interrupted speech is discarded. • While the agent is thinking: there is no distinct behavior documented for false interruptions in the thinking state. Details — Speaking state 1. On detection of voice activity that yields an empty transcription (a false interruption), the agent immediately stops speaking. 2. How it continues depends on two interruption settings: ◦ 
resume_false_interruption
(boolean) ▪︎ True: after a period of silence defined by 
false_interruption_timeout
, the agent resumes speaking from the point it stopped. ▪︎ False: the agent does not resume; the interrupted speech is discarded. ◦ 
false_interruption_timeout
(time) ▪︎ This is the silence window the agent waits before concluding the interruption was false and (if resume is enabled) continuing. Sample defaults in SDK examples use a 2 second timeout. 1. This behavior treats brief non-verbal sounds as a pause rather than a full turn change, keeping conversational flow natural. Thinking state • The documentation does not specify a separate, defined behavior for false interruptions that occur while the agent is in the thinking state. • If you need a predictable action in that state (for example, cancel thinking, restart generation, or log the interruption), you can let me know the behavior you want and I can suggest approaches or configuration options to approximate it. Next steps / offer to help • Do you want the agent to behave in a particular way when interrupted while thinking (e.g., abort thinking and start listening, ignore brief sounds, or emit an event)? Tell me which outcome you prefer and I’ll recommend a concrete setup or code-snippet to implement it. • If you’re seeing behavior that doesn’t match the above, please share the 
turn_handling.interruption
settings you’re using and whether you observed this during speaking or thinking. I’ll help diagnose. Sources: Turns overview | LiveKit Documentation | Adaptive interruption handling | LiveKit Documentation, with interactive elements
b
• While the agent is speaking: the agent immediately stops on detected voice activity. If 
resume_false_interruption
is true, the agent will resume from where it left off after silence for `false_interruption_timeout`; if false, the interrupted speech is discarded. • what do you mean "if its false the the interrupted speech is discarded."
t
Thanks — good question. I’ll clarify what “the interrupted speech is discarded” means. If 
resume_false_interruption
is set to 
false
, when a false interruption (e.g., a cough or “mm‑hm” that produces no transcription) occurs while the agent is speaking: • The agent immediately stops speaking at the moment the sound is detected. • The speech output that the agent had been producing up to that stop is not continued or replayed later — it’s thrown away. In other words, the partial utterance is not resumed or finished automatically. • After stopping, the agent treats the interruption as the end of its turn and will wait for the next user input (or whatever your session logic does on turn end) before producing any new speech. Contrast that with 
resume_false_interruption: true
, where the agent would wait for the 
false_interruption_timeout
silence window and then continue speaking from the point it left off. If you want, I can show a short code example of how to toggle 
resume_false_interruption
and 
false_interruption_timeout
in your AgentSession config, or suggest behavior (e.g., emit an event or restart the response) you can implement when 
resume_false_interruption
is false. Which would you prefer? , with interactive elements
b
in the tutorial video i saw, the agent does not even stop speaking upon, false interrupting with a cough, there is not even a second of agent stopping its speech
t
Thanks — that’s a helpful observation. Let’s clarify why what you saw in the tutorial video might differ from the documented behavior and how to check the cause. Why the agent in the video may not have stopped speaking • The ML model can classify short non-verbal sounds (like a cough or “mm‑hmm”) as non-interruptive — in that case the agent won’t stop because it never treated the sound as an interruption. • The interruption feature may have been disabled or set differently in that demo (for example, 
turn_handling.interruption
turned off). • 
false_interruption_timeout
could be configured extremely short in the demo so any brief pause was below the audible threshold and looked uninterrupted. • The demo might be using a custom flow where the agent mutes incoming VAD events while speaking, or it could be playing pre-recorded audio where the playback isn’t interrupted by input. What “not stopping” means vs the documented behavior • Documented default behavior: when voice activity that looks like an interruption is detected while speaking, the agent normally stops immediately. If 
resume_false_interruption: true
, it will resume after 
false_interruption_timeout
of silence; if false, it won’t resume and current utterance is discarded. • If the agent “does not stop,” that means the system did not detect the input as an interruption (it was classified as non-interruptive) or the interruption handling was effectively disabled/modified in that demo environment. What I can help you check right now • Could you share your 
turn_handling.interruption
settings (values for 
resume_false_interruption
and 
false_interruption_timeout
)? • Was the demo using live microphone input or pre-recorded playback? (Pre-recorded playback can behave differently in demos.) • Are you seeing this behavior locally with your configuration, or only in the tutorial video? If you want a quick test suggestion 1. Enable interruption handling with a clearly non-verbal sound (cough) while the agent speaks. 2. Try both 
resume_false_interruption: true
and 
false
with a visible 
false_interruption_timeout
(e.g., 2s). 3. Observe whether the agent stops and whether it resumes after the timeout. Tell me what you tried or share your 
turn_handling.interruption
config and whether the input was live or playback, and I’ll help pinpoint why the demo behaved differently and recommend settings to reproduce (or suppress) that behavior. , with interactive elements
4 Views
Open in Slack
PreviousNext
Powered by