Telephony and SIP

Why it matters

This documentation explains how developers connect ordinary telephone calls to OpenAI's GPT-Live service using SIP trunks, with providers such as Twilio handling the phone network side. The design splits responsibilities cleanly: the provider exchanges call audio with OpenAI, while the developer's backend manages session configuration, call controls, and business logic. Inbound calls arrive as a webhook event that the backend must accept or reject, with the first decision taking precedence and later competing decisions rejected. Outbound calls skip the webhook flow entirely and are placed by creating a session configured for SIP transport. Security is treated throughout: signaling runs over TLS, audio must use SRTP, webhook signatures require verification, and SIP headers are explicitly labeled as untrusted caller metadata rather than authorization. The piece also addresses practical integration concerns, including deduplicating webhook deliveries, migrating from a deprecated event name, and assigning a single handler to each action so that duplicate events cannot trigger tools twice. A sideband WebSocket carries transcripts, tools, and commands while SIP carries the audio.