Skip to navigation

Realtime event reference

Reference of the events received from the Realtime Speech to Text API

Reference · Lists all events sent to and received from the Realtime Speech to Text WebSocket API.

Review the API reference for the Realtime Speech to Text API for more information on the API and its options.

committed_transcript is the stable, final event for each segment. The public API does not emit a separate final_transcript event.

Sent events

EventDescriptionWhen to use
input_audio_chunkSend audio data for transcriptionContinuously while streaming audio

With Scribe v3 Live, the first input_audio_chunk of a session can also include a keyterm library.

Received events

EventDescriptionWhen received
session_startedConfirms connection and returns session configurationImmediately after WebSocket connection is established
partial_transcriptLive transcript update for the in-progress segment. Later partials replace earlier ones. With diarize=true, includes draft words with speaker_id (null until identified); speaker IDs and timestamps may change, and unaligned words have no timestamps. Also sent as a keepalive — empty text if the segment has no uncommitted text, or a repeat of the latest partial if it doesDuring audio processing, before a commit is made. Also sent as a keepalive about once per keepalive_interval_ms when configured
committed_transcriptSettled transcript for a segment. This is the stable, final text for the segment and will not change. With diarize=true, includes a zero-based segment_index and is not delayed by diarizationAfter a commit (either manual or VAD triggered)
diarized_transcriptSpeaker-labeled transcript with text, words and a segment_index matching the committed transcript. Unresolved speakers have speaker_id: null. See speaker diarizationAfter every committed transcript, in commit order, even if speakers cannot be identified. Only received when diarize=true
speakers_mergedIndicates that from_speaker_id and to_speaker_id refer to one person. Replace every earlier use of from_speaker_id in partial and diarized transcripts with to_speaker_id; apply chained merges in receive order. IDs are never reusedWhen two speaker IDs are identified as the same person. Only received when diarize=true
committed_transcript_with_timestampsCommitted transcript with word-level timestamps and the detected language_codeSent after the committed transcript. Only received when include_timestamps=true or include_language_detection=true is included in the query parameters
committed_transcript_entitiesEntities detected in the committed transcript segment, with character offsetsSent shortly after each committed transcript. Only received when entity_detection is included in the query parameters
redacted_transcriptCommitted transcript segment with the redacted entities replaced by markers, alongside the original committed textSent shortly after each committed transcript. Only received when entity_redaction is included in the query parameters
edited_transcriptEdited version of the committed transcript segment, alongside the original committed textSent shortly after each committed transcript. Only received when transcript_edit is included in the query parameters

Language code

Set include_language_detection=true in the query parameters to receive the language of the speech on partial_transcript and committed_transcript_with_timestamps events as language_code.

  • On partial_transcript, language_code is always present and is null until the in-progress segment has enough speech to tell the language (about four words, or eight characters for languages written without spaces, such as Chinese, Japanese and Thai) and the same language has been seen on consecutive partials. Very short segments, such as a single word, may stay null, and a partial too short to tell resets the count of consecutive partials. Once set, it only changes after the new language has been seen on consecutive partials, so a single odd partial does not flip it.
  • Partials with no spoken text, such as only punctuation, have language_code: null.
  • language_code on committed_transcript_with_timestamps is the final answer for the segment. If it differs from the last partial, use the committed value. It is sent after committed_transcript, so show the partial’s value until it arrives.
  • If you pass a language_code or secondary_languages when connecting, the language comes from the model’s recognition within those languages. Without a hint, it is detected from the transcript text.
  • Without include_language_detection=true, partial_transcript events have no language_code field.

Error handling

If an error occurs, an error message will be returned before the WebSocket connection is closed.

Error TypeDescription
auth_errorAn error occurred while authenticating the request. Double check your API key
quota_exceededYou have exceeded your usage quota
transcriber_errorAn error occurred while transcribing the audio.
input_errorAn error occurred while processing the audio chunk. Likely due to invalid input format or parameters
invalid_requestThe connection parameters were rejected. Check the query parameters against the API reference
errorA generic server error. Also sent when entity redaction fails, before the connection is closed
commit_throttledThe commit was throttled due to too many commit requests made in a short period of time
unaccepted_termsYou haven’t accepted the terms of service to use Scribe. Please review and accept the terms & conditions in the ElevenLabs dashboard
rate_limitedYou are rate limited. Please reduce the amount of requests made in a short period of time
queue_overflowThe processing queue is full. Please send fewer requests made in a short period of time
resource_exhaustedServer resources are at capacity. Please try again later
session_time_limit_exceededMaximum session time has been reached. Please start a new session or upgrade your subscription
chunk_size_exceededThe audio chunk size is too large. Please reduce the chunk size
insufficient_audio_activityYou haven’t sent enough audio activity to maintain the connection