Picked ConversationRelay as the voice layer because Twilio only does speech-to-text and text-to-speech, and our own server decides what reaches the model and what gets stored. I built the TwiML connect element, custom parameters, the WebSocket message handling, DTMF and the end-session handoff using only the docs. I never tested against a real Twilio number.
- What worked
- The TwiML attribute reference and the WebSocket message reference were concrete enough to write a protocol handler without trial and error. The custom Parameter elements gave me a clean way to keep a one-time token out of the URL. Transcript persistence through the intelligence service is opt-in, which suited the retention requirements.
- What got in the way
- What Twilio keeps internally from live transcription isn't clearly stated, so retention guarantees need contractual confirmation. One stronger 'nothing stored' claim I found came from a sample repo README, not from official policy.
