What changed
OpenAI's official Python SDK v3.34.0, released on October 17, adds Realtime Translation WebSocket sessions. The release connects the SDK to OpenAI's dedicated gpt-realtime-translate model and its translation-specific realtime endpoint.
The important shift is that this is not a general voice agent being prompted to translate. A translation session continuously accepts source audio and streams back translated audio plus transcript deltas while the speaker is still talking.
Why the architecture matters
The Realtime Translation guide separates an interpreter from a conversational assistant: translation sessions use /v1/realtime/translations, start from the incoming audio stream, and do not use the turn-based response.create lifecycle. That makes the primitive a better fit for media products that already have an audio path.
OpenAI documents two transports. WebRTC is intended for browser capture and playback; WebSockets fit servers that already receive raw audio from sources such as Twilio Media Streams, SIP, broadcasts, or media workers. The model currently supports more than 70 input languages and 13 target output languages, which is enough coverage for a useful first version of multilingual calls, events, and learning products.
What builders can ship
This turns translation into a sidecar that can be attached to existing live experiences:
- A conference or livestream can publish translated audio and captions without rebuilding its player.
- A support or sales call can create one translation session per direction and let each participant choose a target language.
- A video room can add a listener-specific translation layer while keeping the original media track intact.
The model documentation and the official cookbook walkthrough show the practical flow: stream continuous audio in, receive translated audio chunks and transcript events, and keep the standard API key on the server. Browser apps should mint short-lived client secrets rather than expose that key to the client.
The product takeaway
The useful mental model is source audio -> translation session -> translated audio + subtitles, not chatbot with a translation prompt. That separation gives product teams a clean place to measure the two things users actually notice: meaning and delay. Before shipping, test with real audio, bilingual review, names and numbers, overlapping speakers, and the latency from source speech to translated audio.
For AI product builders, the opportunity is less about adding another model selector and more about making multilingual participation a toggle inside an existing call, broadcast, or classroom workflow.