Feature Type
Nice to have
Feature Description
Problem
We are running LiveKit Agents in a fully self-hosted environment for SIP/telephony voice agents.
Our users are primarily Hindi and Telugu speakers. During calls, users frequently provide conversational backchannels such as:
haa, avunu,hmm,yeah, okay
These are acknowledgements rather than genuine interruptions.
With VAD-based interruption handling, these utterances often trigger unwanted agent interruptions, creating a poor conversational experience.
Current Limitation
Adaptive Interruption Handling works well but relies on LiveKit-hosted inference. Self-hosted deployments eventually fall back to VAD-based interruption handling.
For organizations that require fully self-hosted deployments, there is currently no equivalent local interruption model.
Feature Request
Please provide one or more of the following:
- Self-hostable Adaptive Interruption model.
- Official Krisp VIVA Turn Prediction / Interruption Prediction integration.
- Pluggable interruption detector interface allowing custom interruption models.
- Local inference option for adaptive interruption handling
Expected Outcome
Allow self-hosted LiveKit deployments to distinguish:
Backchannels:
from true interruptions:
- wait
- stop
- actually
- let me explain
without requiring LiveKit Cloud inference
Workarounds / Alternatives
No response
Additional Context
No response
Feature Type
Nice to have
Feature Description
Problem
We are running LiveKit Agents in a fully self-hosted environment for SIP/telephony voice agents.
Our users are primarily Hindi and Telugu speakers. During calls, users frequently provide conversational backchannels such as:
haa, avunu,hmm,yeah, okay
These are acknowledgements rather than genuine interruptions.
With VAD-based interruption handling, these utterances often trigger unwanted agent interruptions, creating a poor conversational experience.
Current Limitation
Adaptive Interruption Handling works well but relies on LiveKit-hosted inference. Self-hosted deployments eventually fall back to VAD-based interruption handling.
For organizations that require fully self-hosted deployments, there is currently no equivalent local interruption model.
Feature Request
Please provide one or more of the following:
Expected Outcome
Allow self-hosted LiveKit deployments to distinguish:
Backchannels:
from true interruptions:
without requiring LiveKit Cloud inference
Workarounds / Alternatives
No response
Additional Context
No response