Skip to content

Rumble-2-Flash is now open weights, Apache 2.0.

Read the announcement

Diarization

Speaker diarization partitions transcribed speech by speaker - answering “who spoke when”. When enabled, each piece of transcribed text is labelled with a speaker_id so you can tell participants apart in a conversation.

Diarization is available on the Prerecorded and Realtime APIs. It is not available on the Turns API.

Set the diarize query parameter to true. Optionally cap the number of distinct speakers with max_speakers (1-8); leave it unset to let the server determine the count automatically.

?diarize=true&max_speakers=2

Diarization distinguishes up to 8 speakers per session (prerecorded request or realtime connection).

  • max_speakers accepts values from 1 to 8. Any other value is rejected with 400 Bad Request.
  • max_speakers is an upper bound, not an expected count. If you know a call has two participants, max_speakers=2 keeps their speech from being split across extra speaker IDs; fewer IDs may still appear if fewer people speak.
  • Audio with more than 8 distinct speakers can still be transcribed, but some speakers will share a speaker_id, so IDs no longer map one-to-one to people.

Each speaker is identified by an integer speaker_id from 0 to 7 (or up to max_speakers - 1 when set). IDs are assigned in the order speakers are first heard and remain stable for the duration of a session.

On prerecorded, the response is split into per-speaker segments, each with its own speaker_id and text. On realtime, each transcript message carries the speaker_id of its dominant speaker. See the API reference for Prerecorded and Realtime for the full response fields.