litellm - 💡(How to fix) Fix ElevenLabs speech-to-text via LiteLLM does not appear to pass diarization options

Summary

When calling ElevenLabs Speech-to-Text through LiteLLM, diarization-related request parameters do not appear to be honored/passed through to the upstream ElevenLabs /v1/speech-to-text API. This makes transcriptions come back without the expected speaker separation even when diarize=true and related ElevenLabs parameters are included.

Expected behavior

LiteLLM should forward supported ElevenLabs Speech-to-Text multipart form fields to ElevenLabs, including at least:

• diarize
• tag_audio_events
• timestamps_granularity
• num_speakers
• diarization_threshold
• detect_speaker_roles
• language_code

ElevenLabs documents these fields for POST https://api.elevenlabs.io/v1/speech-to-text with model_id=scribe_v2.

Docs: https://elevenlabs.io/docs/api-reference/speech-to-text/convert

Actual behavior

Calling the model through LiteLLM produced transcript output that did not include the expected diarization/speaker separation, even though the request included diarization settings. The same integration is being changed to optionally call ElevenLabs directly because the LiteLLM gateway behavior appears to be the source of the missing diarization.

Example request shape

The application sends a multipart request to LiteLLM transcription with fields like:

model=elevenlabs/scribe_v2
response_format=verbose_json
timestamp_granularities[]=segment
timestamps_granularity=word
diarize=true
tag_audio_events=true
detect_speaker_roles=true

Optional fields that should also pass through when present:

num_speakers=<int>
diarization_threshold=<float>
language=<code>
language_code=<code>

Environment/context

• LiteLLM gateway model: elevenlabs/scribe_v2
• ElevenLabs direct endpoint expected: POST /v1/speech-to-text
• ElevenLabs model id: scribe_v2

Request

Please confirm whether LiteLLM's ElevenLabs Speech-to-Text provider currently forwards these extra multipart parameters. If not, please add pass-through support for the documented ElevenLabs transcription options, especially diarize, timestamps_granularity, and detect_speaker_roles.