Gpt-realtime Model with_azure function - API version 2025-08-08

[michael-wimmer-mw]

Bug Description

Hi,

we currently try to migrate the api version for the azure gpt realtime model to version 2025-08-08 as the 2024-10-01-preview is announced to be retired on end of april 2026.
From the documentation
RealtimeModel documentation

I can see that objects for input_audio_transcription have changed from
openai.types.beta.realtime.session.InputAudioTranscription (API-Version 2024-10-01)
to
openai.types.realtime.AudioTranscription (API-Version 2025-08-08).

However, in the code I can see that only the object InputAudioTranscription is accepted by the with_azure function

Github realtime_model.py

A similar behaviour can be found for turn_detection.
The model cannot be instantiated with the new objects but api-version 2025-08-08 does not work with the old objects as well.

Is there a fix planned or are there any known workarounds for this?

Thank you

Expected Behavior

Realtime Model can be instantiated with api version 2025-08-08 as described in the documentation

Reproduction Steps

1. 
2.
3.
...
- Sample code snippet, or a GitHub Gist link -

Operating System

Unix

Models Used

No response

Package Versions

livekit-agents==1.3.12

Session/Room/Call IDs

No response

Proposed Solution

Additional Context

No response

Screenshots and Recordings

No response

[dhansmair]

any updates on this?

[Nivg]

I tried to work around it with this:

def create_realtime_model() -> openai.realtime.RealtimeModel:
    endpoint = os.environ["AZURE_OPENAI_ENDPOINT"].rstrip("/")
    deployment = os.environ["AZURE_OPENAI_DEPLOYMENT"]
    return openai.realtime.RealtimeModel(
        model=deployment,
        base_url=f"{endpoint}/openai/v1/realtime?model={deployment}",
        api_key=os.environ["AZURE_OPENAI_API_KEY"],
        api_version="",
        input_audio_transcription=AudioTranscription(model="whisper-1"),
        voice="alloy",
    )

Audio works, but there's no transcript and no live video from screen share — so no vision (which I definitely need). This is using the latest model version: 2025-08-08. My conclusion is that Azure OpenAI Realtime API support with LiveKit just isn't there yet.

Here's some Azure reference in case anyone finds it useful: GPT Realtime API for speech and audio