Interruption Prediction Now Available for Telnyx Voice AI Assistants

Voice AI Assistants on Telnyx can now tune how confident they must be that a caller wants to take the turn before they stop talking. Set a threshold and the assistant ignores backchannels like "yeah" or "uh-huh", only yielding the floor for genuine interruptions.

What's new

• Interrupt prediction threshold: A new slider under Speaking Plan, interruption_settings.interrupt_prediction_threshold, sets the confidence required before barge-in fires. Range 0.0 to 1.0, default 0.0 (off), 0.4 is a good starting point.
• Backchannel suppression: Below the threshold, short acknowledgments no longer trigger the assistant to stop mid-utterance, so the assistant keeps its turn instead of stuttering every time the caller nods along.
• Per-assistant config: Each assistant carries its own threshold in interruption_settings, configurable from Mission Control or the AI Assistants API. Null or 0.0 disables prediction and falls back to standard barge-in.
• Deepgram Flux integration: Available when Deepgram Flux is the transcription provider, across English, German, Spanish, Italian, and Japanese.

Why it matters

• False interruptions are one of the most common reasons voice AI calls feel broken. The assistant stops, restarts, and the caller has to repeat themselves, which tanks completion rates on long calls.
• Backchannels are conversational glue, not turn requests. A model that treats "uh-huh" as "stop talking" will fragment every longer answer into a half-finished sentence followed by an awkward restart.
• On-network processing keeps the prediction inside Telnyx's voice pipeline, so the barge-in decision is made in the same system that's handling the audio, with no extra hop to a third-party endpoint.
• Tunability matters more than defaults. Support agents, sales agents, and dispatch agents each want a different sensitivity, and a single global on/off switch would force every workload into one pacing.

Example use cases

• Support agents that read longer scripted answers can keep talking through a customer's "yeah, okay" without stopping and restarting the same sentence.
• Sales agents running discovery can stay in their turn during a prospect's verbal nods and only yield when the prospect actually starts a question.
• Multilingual contact centers can apply the same threshold across English, German, Spanish, Italian, and Japanese lines without per-language tuning.
• Voice apps with high backchannel cultures, where callers acknowledge frequently, can stop treating every acknowledgment as an interruption without disabling barge-in entirely.

Getting started

1. In Mission Control, go to AI → Assistants and open the assistant you want to configure.
2. Select the Voice tab and expand the Speaking Plan section.
3. Make sure Interruptions is set to Enabled.
4. Set Deepgram Flux as your transcription model under Transcription.
5. Drag the Interruption Prediction Threshold slider to your target value. Start at 0.4 and adjust up for stricter gating or down for more permissive barge-in. Set to 0 to disable.

Learn more in the Telnyx transcription settings docs.