Twilio Changelog | May. 06, 2026
Conversation Relay now supports Deepgram Flux + New Features
Enhance your voice agents with Deepgram’s Flux model for better turn-taking and fewer interruptions. This update also introduces new TwiML attributes for precise conversation flow control and a dedicated suite of error codes for easier debugging.
Deepgram Flux STT Model Support
Conversation Relay now supports Deepgram's latest speech recognition model, Flux, bringing fast, natural turn detection to your voice agents including multilingual conversations across ten languages. Flux fuses transcription and turn detection in a single model, eliminating the race conditions that cause agents to cut callers off mid-sentence or wait too long to respond. The result: up to 200-600ms less agent response latency and ~30% fewer false interruptions (learn more from Deepgram here).
New TwiML Attributes
Three new attributes give you more control over conversation flow.
speechTimeout sets how long ConversationRelay waits after speech ends before finalizing a prompt, letting you tune for the pace of your callers.
ignoreBackchannel filters out conversational filler — "yeah," "uh-huh," "okay" — so these words don't interrupt your agent's TTS playback; detection covers English, Spanish, French, German, and more.
The new events attribute lets you subscribe to session events like speaker-events and tokens-played, replacing the previous debug attribute options (which remain supported for backward compatibility).
Error Codes
A new error code reference documents twelve Conversation Relay-specific codes (64101–64112) covering invalid parameters, WebSocket failures, concurrency limits, TTS errors, and more — including whether each condition ends the session.
Check out the blog post to learn more. If you’re ready to dive right into code you can find the docs here. We can’t wait to see what you build!