Skip to main content

Interface: OpenAIRealtimeTranscriptionRollover

Defined in: packages/agentos/src/io/voice-pipeline/providers/OpenAIRealtimeTranscriptionSTT.ts:157

When and how a session moves to a new connection before OpenAI's session limit (60 minutes for a Realtime session). Ages are counted from the moment a connection's session update was confirmed.

With server turn detection, a connection older than afterMs opens the next one at its first end of speech, or at deadlineMs whatever is being said. Both connections receive the same audio for at least overlapMs; the old one keeps receiving audio until it is between utterances (or until hardStopMs, when its buffer is committed), finishes the items it holds and closes. An utterance both connections transcribed is emitted once: from the old connection, or from the new one when the old connection's transcription of it fails, comes back empty or never arrives. The new connection's interim text for utterances that began while the old one was open is not emitted. With client commits (turnDetection: null) the switch happens at the first flush() after afterMs, or with a commit at deadlineMs, and the connections do not overlap.

Properties​

afterMs?​

optional afterMs: number

Defined in: packages/agentos/src/io/voice-pipeline/providers/OpenAIRealtimeTranscriptionSTT.ts:159

Connection age after which the first end of speech opens the next connection.

Default Value​

3_300_000 (55 minutes)

approve()?​

optional approve: (request) => boolean | Promise<boolean>

Defined in: packages/agentos/src/io/voice-pipeline/providers/OpenAIRealtimeTranscriptionSTT.ts:184

Called before a rollover opens the next connection. Resolve false, or throw, to refuse it: the session keeps its current connection until hardStopMs, finishes the items it holds, and closes. An answer that has not come by hardStopMs counts as a refusal, and a later one opens nothing. It is a policy check beside the funding: a host that pays for audio per connection reserves each one, rollovers included, through StreamingSTTConfig.connectionHooks.authorize, keyed by the connection's connectionId rather than this numeric index.

Parameters​

request​

OpenAIRealtimeRolloverRequest

Returns​

boolean | Promise<boolean>

deadlineMs?​

optional deadlineMs: number

Defined in: packages/agentos/src/io/voice-pipeline/providers/OpenAIRealtimeTranscriptionSTT.ts:161

Connection age at which the next connection opens whatever is being said.

Default Value​

3_480_000 (58 minutes)

hardStopMs?​

optional hardStopMs: number

Defined in: packages/agentos/src/io/voice-pipeline/providers/OpenAIRealtimeTranscriptionSTT.ts:171

Connection age at which an old connection still in speech is committed and drained, and at which a connection whose rollover was refused, or has not opened its next connection yet, ends the session. Keep it under the provider's 60-minute limit.

Default Value​

3_570_000 (59.5 minutes)

matchToleranceMs?​

optional matchToleranceMs: number

Defined in: packages/agentos/src/io/voice-pipeline/providers/OpenAIRealtimeTranscriptionSTT.ts:173

Slack, in milliseconds, when matching an utterance of the new connection to one of the old.

Default Value​

500

overlapMs?​

optional overlapMs: number

Defined in: packages/agentos/src/io/voice-pipeline/providers/OpenAIRealtimeTranscriptionSTT.ts:163

Least time both connections receive the same audio.

Default Value​

3_000