The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 18, 2026 2:07am

Greptile Summary

This PR fixes a data-loss bug where Mistral Voxtral's diarization fields (segments and language) were silently dropped during transform_audio_transcription_response, leaving users unable to access speaker-attributed transcript segments.

Changes:

• litellm/llms/mistral/audio_transcription/transformation.py: After constructing the base TranscriptionResponse, the fix conditionally copies segments and language from the raw API response JSON to the top-level response object using dictionary-style attribute assignment (response["key"] = value).
• tests/test_litellm/llms/mistral/audio_transcription/test_mistral_audio_transcription_transformation.py: Adds test_mistral_audio_transcription_response_transform_diarized, a fully mocked unit test verifying that both segments (with speaker_id, start, end) and language (including None) survive the transformation.

The implementation correctly follows the pattern already used by Deepgram (response["language"], response["words"]) and ElevenLabs (response["language"]), and no real network calls are made in the new test.

Confidence Score: 5/5

• This PR is safe to merge — it is a small, additive, non-breaking change with appropriate mock test coverage.
• The change is minimal (7 lines), follows an already-established pattern used by Deepgram and ElevenLabs, introduces no breaking changes (segments/language are only surfaced when present in the response), and is fully covered by a mock unit test. No hardcoded model flags, no FastAPI imports, no DB calls, and no backwards-incompatible behaviour changes.
• No files require special attention.

Sequence Diagram

sequenceDiagram
    participant Client
    participant LiteLLM
    participant MistralAPI

    Client->>LiteLLM: transcription(model="mistral/voxtral-mini-latest", diarize=True)
    LiteLLM->>MistralAPI: POST /v1/audio/transcriptions (form: diarize=true)
    MistralAPI-->>LiteLLM: { text, language, segments: [{speaker_id, start, end, ...}], usage }
    Note over LiteLLM: transform_audio_transcription_response()<br/>Extract text → TranscriptionResponse<br/>Preserve segments → response["segments"]<br/>Preserve language → response["language"]<br/>Store full JSON → _hidden_params
    LiteLLM-->>Client: TranscriptionResponse(text, segments, language)

_{Last reviewed commit: "fix(mistral): preser..."}