GPT Realtime Translate

GPT-Realtime-Translate is a specialized OpenAI model for streaming speech-to-speech translation, released as part of the Realtime API family. It ingests source audio and returns translated audio plus transcript deltas while the speaker is still talking, supporting both source and target transcripts.

Translation sessions use a dedicated /v1/realtime/translations endpoint rather than the standard /v1/realtime endpoint used by voice agents. The architecture is built around continuous audio streaming (no response.create calls), with the model acting as an interpreter rather than an assistant. Both WebRTC (media track transport) and WebSocket (base64-encoded 24 kHz PCM16) connections are supported.

Typical use cases include simultaneous interpretation, multilingual broadcasts, meetings, lessons, conference calls, and customer support. Pricing is based on audio duration (USD per minute) rather than tokens.