TwilioJesse Sumrak12 min readintermediate
How to handle real-time interruptions in your AI voice agent
Summary
Twilio’s blog explains how to manage barge‑in, backchannels, and noisy speech in AI voice agents using Conversation Relay settings and Deepgram Flux turn‑detection, and shows how to keep LLM context in sync after an interruption.
- Configure Conversation Relay attributes (interruptible, ignoreBackchannel, eotThreshold, speechTimeout) to handle barge‑in, backchannels, and background noise without writing new code.
- Deepgram Flux fuses transcription and turn detection, reducing false interruptions by ~30% and cutting latency by 200‑600 ms compared to Nova‑3.
- When an interruption occurs, use the WebSocket interrupt message (utteranceUntilInterrupt) to trim the LLM’s conversation history so the model only sees what the caller actually heard.
- Adjust interruptSensitivity and eotThreshold (0.5‑0.9) to balance responsiveness against false positives in noisy environments.
Voice‑AI engineers and contact‑center developers need these knobs to make agents robust in real‑world, noisy calls and to keep LLM responses coherent after interruptions.
6/10
