Twilio ConversationRelay

ConversationRelay is a Twilio product that connects incoming and outgoing phone calls to a developer's own large language model or conversational AI logic. It sits in the middle of the call, handling the telephony plumbing: speech-to-text transcription, text-to-speech playback, audio streaming over WebSockets, and detecting when a caller interrupts (barge-in) so the agent can stop talking and listen. Developers write a WebSocket server that receives transcribed text and DTMF events from Twilio and sends back text or SSML for the AI to speak, plugging in whatever LLM or orchestration layer they choose rather than being locked into a single model provider.

It's built for teams building voice agents for customer support, appointment scheduling, order status, and similar phone-based automation who don't want to stitch together separate STT, TTS, and telephony infrastructure themselves. Twilio handles things like automatic language detection, low-latency turn-taking, and the option to transfer a call to a live human agent mid-conversation when the AI can't resolve an issue. It also integrates with Twilio's Conversational Intelligence product for post-call analytics, transcript storage, and things like hallucination or escalation detection.

What sets it apart from typical voice AI platforms is that it's infrastructure rather than a full agent framework: Twilio isn't selling the conversational brain, just the reliable pipe between phone calls and whatever model or agent logic a developer runs on their own backend. That makes it attractive to teams who already have Twilio numbers and call flows and want to add AI without rebuilding their telephony stack, though it does mean more assembly work compared to fully managed voice agent platforms.