Speaker Attributed ASR (SAA) is a rich transcription mode that asks the ASR model itself (Granite Speech 4.1 Plus or Nemotron 3.5 Transcribe) to produce a transcript with speaker-turn labels. Use it when you want a speaker-attributed text transcript, not acoustic diarization segments. For “who spoke when” timelines with timestamps, use Diarization.

When to Use SAA


Model Requirement

SAA requires one of:
If the API request omits model_id, the persisted SAA workflow defaults to Granite-Speech-4.1-2B-Plus. Supplying any other model returns a validation error. Nemotron-3.5-Transcribe-30B-A4B-FP8 needs about 32 GiB of memory (see Transcription). It tags speakers as it transcribes, in one pass of up to 60 minutes. Its turns are labelled “Speaker 1”, “Speaker 2” and so on, and carry start and end times when its own word timestamps align. It labels speakers in Arabic, Chinese, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Spanish and Thai. The speaker expectation does not steer it: a speaker count outside min_speakers or max_speakers only adds a warning to the record.

Using the Web UI

  1. Open Transcription in the sidebar.
  2. Choose Speaker Attributed ASR from the mode switch.
  3. Upload or record audio.
  4. Select a ready speaker-attributed model.
  5. Choose a speaker expectation: Auto, 2+, 3+, or 4+.
  6. Optionally enable summary generation.
  7. Submit the job and review the speaker-turn transcript.
SAA disables streaming and timestamp alignment. Those controls belong to normal transcription or diarization workflows.

Using the API

Create a persisted SAA job with job_kind=speaker_attributed_asr:
The short alias job_kind=saa is also accepted. Poll the returned record until processing_status is ready:
Fetch stored audio or regenerate summaries with the same job kind:

Request Fields

JSON and multipart create requests accept: Unsupported for SAA:
  • stream
  • include_timestamps
  • word_timestamps
  • aligner_model_id

See Also