Transcript Lifecycle

PreviousNext

StableStream™ output and LiveFinal™ processing in the public Realtime API.

This page describes the Lanson Live API public Realtime endpoint, WS /v1/audio/transcriptions/stream. LiveFinal™ names the correction/finalization mechanism; StableStream™ names the stable, readable output behavior. Neither name is an event type.

State flow

speech_started → speech_stopped → recognition → optional refinement → completed
                                                                  → optional translation

Recognition and refinement are server processing stages. The current endpoint exposes a single completed transcript per successfully processed utterance, not partial → revision → final events.

Event semantics

speech_started

VAD detected the start of a speech segment.

  • At this point: no transcribed text yet
  • Client action: optionally show a "listening" indicator
  • Mutable: no, utterance_index is fixed

speech_stopped

VAD detected the end of a speech segment (silence or a segmentation boundary).

  • At this point: recognition begins; optional refinement follows before delivery
  • Client action: optionally show "processing" state
  • Mutable: no

conversation.item.input_audio_transcription.completed

A speech segment has been transcribed. This is the primary output.

  • At this point: text is stable and will not change
  • Client action: display or persist text, ordered by utterance_index
  • Mutable: no, text is final
  • Can trigger downstream: check transcript.safe_for_downstream and your application's accuracy requirements

Correction and translation

Enable refinement before audio when the gateway reports correction availability. It is off by default; the default lite profile is included in the base transcription price, and pro is billed as the realtime refinement add-on. Correction runs before completed; raw_text preserves recognition output, text is the delivered result, and transcript.correction_applied reports whether correction changed it. If correction fails, the service falls back to the raw output.

transcript.state is currently stable and transcript.revision is 0. The public endpoint does not send a separate correction event, later transcript revisions, or a session-wide final event.

When translation is enabled, live.translation.completed arrives independently for each target language. It uses raw source text and associates the result by utterance_index. A missing translation should not block rendering the source text.

utterance_index

Each speech segment has a unique, incrementing utterance_index. Use it to:

  • Order transcribed results
  • Detect missing events (index gaps)
  • Merge into a complete transcript

Concurrent utterances

The upstream may process multiple speech segments simultaneously. limits.max_concurrent_utterances limits concurrency. When exceeded, frames are dropped and lanson.throttled is sent.

No partial text

In the current version:

  • No partial transcription is pushed: no intermediate text before transcription.completed
  • One result per utterance: each utterance produces one completed event
  • Text is immediately stable: no partial → final transition needed on the client

This means clients do not need a partial-text replacement protocol on this endpoint. They still need to order concurrent results, handle language uncertainty and errors, and avoid layout shifts in their UI.

Persist completed results as they arrive. Do not infer that a flush message or socket close will produce a full-session final transcript; see Client Messages.