LiveFinal™ is LansonAI's processing mechanism for progressive correction and segment finalization. It describes how speech moves toward a usable result as processing and context become available; it is not a selectable model, request flag, or WebSocket event name.
Trinity Engine coordinates the processing. LiveFinal™ describes correction and finalization. StableStream™ describes the stable, readable text experience delivered to an application.
What you receive today
For the public WS /v1/audio/transcriptions/stream endpoint, the successful per-utterance path is:
speech_started → speech_stopped → recognition → optional refinement
→ transcription.completed
→ optional translation.completedThe abbreviated names above refer to the exact events in the table below. Recognition and refinement are processing stages, not additional client events.
| Stage | Public event or field | Application behavior |
|---|---|---|
| Speech begins | input_audio_buffer.speech_started | Show a listening indicator; there is no transcript yet. |
| Speech segment ends | input_audio_buffer.speech_stopped | Show processing while the segment is recognized. |
| Optional correction | session.update.refinement enables it before audio, when available | Correction runs before the completed result; it is off by default. |
| Segment delivered | conversation.item.input_audio_transcription.completed | Display text, preserve raw_text if needed, and use utterance_index to order results. |
| Delivery state | transcript.state, transcript.revision, transcript.correction_applied | Current state is stable, revision is 0, and the correction flag tells you whether the text changed. |
| Optional translation | live.translation.completed | Attach each target-language result to the same utterance_index. |
Check session.created.features.asr_correction for availability, opt in with refinement: true before sending audio, and confirm session.updated.features.refinement. The default lite profile is included in the base transcription price; pro is billed as the realtime refinement add-on. If unavailable, the server returns refinement_unavailable. See Session Configuration and Pricing.
raw_text preserves the recognition output. text contains the delivered result after optional refinement. If refinement fails, the service falls back to the recognition output. Translation is delivered independently and is based on the raw source text on this endpoint.
Progressive processing is not a partial-text protocol
The current public endpoint emits one completed result per successfully processed utterance. It does not expose a partial → corrected → final sequence, and it does not send a separate live.correction.completed event or a session-wide final transcript.
Do not invent livefinal, session_transcript, or corrected_transcription message types, or interpret revision: 0 as a promise that later revisions will arrive. Product experiences and other service endpoints can expose different stages; their behavior is not automatically part of this public API contract.
The stream uses server-side speech segmentation. A flush message or closing the socket is not a documented guarantee of final transcript delivery. Keep receiving until your expected completed results arrive, and persist them as they are delivered.
Stable does not mean infallible
stable means that this endpoint will not replace that delivered segment through a later revision event. Recognition can still be wrong. safe_for_downstream is false for empty text or a language-detection error; true is not an accuracy guarantee. Handle provisional languages, errors, missing utterances, and application-specific review separately.
Related
- StableStream™ — readable output behavior
- Trinity Engine — processing foundation
- Transcript Lifecycle — states and client behavior
- Server Events — exact event names and fields