Feature request: incremental external context for Realtime Transcription
The Responses API already has this. And I doubt that it trickles into its own “nested” transcription side-channel, so here it goes:
It would be useful for Realtime Transcription to support a context mechanism similar to the Responses API’s conversation items endpoint of its API.
A client could append text context over time — such as terms from application level context which incrementally emerege over time over the span of the session.
The service could then manage retention, prioritization, compaction, and truncation within the transcription model’s context window, rather than application code having to maintain the single context element of the transcription session, which is brittle to do since you can’t really guess how much of that context is necessary for the transcription’s internal operation.
Conceptually, something like:
transcription_context.append(items)
or a session-level equivalent of:
POST /conversations/{id}/items
This would provide for incrementally enriching transcription context while leaving context-window management at the hands of OpenAI’s cloud backend. The Responses API already enables exactly that, according to its documentation at least, and if it’s proven generally effective for the Responses API it can probably be very effective for transcription as well, perhaps even more than it is for the Responses API.
This request holds whether or not the Realtime transcription API already uses earlier transcriptions as context or not.