Terminology

• A deployment is your tuned speech recognition engine, running and addressable by its deployment id. A deployment answers on one of the two APIs: it is provisioned either for real-time streaming or for synchronous transcription, and that determines which API its id works with.
• A turn is a single spoken segment from the user, processed and transcribed as one unit. It is the engine's unit of work on both APIs.
• Turn partials are incremental results emitted while a turn is still being decoded, containing the transcript understood so far. They exist on the Real-time API only, as turn_partial events.
• A session is the full duration of a WebSocket connection. It begins with start_session and typically contains multiple turns as the user and agent interact back and forth. Sessions exist on the Real-time API only.
• Turn-taking mode refers to how turn boundaries are controlled. The current production API supports manual mode, where the client explicitly marks both the start (via start_turn) and the end (via end_turn) of each turn. Automatic turn-taking is on the roadmap.

How the two APIs use these termsCopy Anchor Link

On a Real-time API deployment, one connection is one session, and you bracket every turn inside it yourself.

On a Sync API deployment, one request is one audio file. There is no session, no turn-taking mode, and no partials: you submit the audio for a turn and the response carries its final transcript. Split a long recording into turns and send one request each, which also gives the engine the per-turn context it is tuned for.

Next: the Real-time session flow for how these fit together message by message, or Contextual biasing for the feature both APIs share.