Transcription using gpt-4o-transcribe with gpt-realtime is failing in useast2

Question

Hello,

I am trying to use gpt-4o-transcribe with gpt-realtime in useast2, and it is consistently failing.
I am using gpt-realtime with websockets as per the documentation.
I am seeing the following event: conversation.item.input_audio_transcription.failed

{
   "code":None,
   "message":"Input transcription failed for item 'item_D81pLGqzIyXJ3JdRHoIhD'.",
   "param":None,
   "type":"server_error"
}

The audio format sent in the session is:

{
   "type":"audio/pcm",
   "rate":24000
}

However, when I swap over to whisper-1, it works fine.

I have look for other reports/questions, and have seen varying reports of either:

• The model being retired — although I can't see anything of the sort in the documentation
• Unstable/intermittent failures with this model.

The current setup is working when pointing to OpenAI's hosted realtime models.
Does anyone know what could be the issue?

Answer 1

Hello PRABU WEERASINGHE,

Thank you for the detailed information, and I completely understand the frustration, especially after experiencing this for over a week.

Based on everything you’ve shared, here are the key clarifications:

Do you need to explicitly deploy gpt-4o-transcribe?

Yes, in Azure OpenAI, all models must be deployed to your resource before use, including gpt-4o-transcribe.

Even if the model appears as “available” in the Azure portal for the region, that only means:

• The model is supported in that region

It is eligible for deployment

It does not mean it is automatically provisioned for your resource.

whisper-1 working without additional action likely means it was already deployed in your resource (or you were referencing an existing deployment).

Please confirm:

Go to Azure OpenAI → Model Deployments

Verify that gpt-4o-transcribe is deployed

Confirm you are referencing the deployment name, not the base model name, in your websocket connection

Azure requires calling the deployment name, not the model name directly.

Why the server_error with no code?

The error:

conversation.item.input_audio_transcription.failed
type: server_error
code: None

strongly indicates one of the following:

The model is not properly deployed

Deployment is in a failed or degraded backend state

The deployment SKU does not support realtime transcription

A regional backend issue specific to useast2

Since PCM 24kHz works with whisper-1,

It works against OpenAI hosted endpoints,

Your websocket implementation is valid

This narrows it down to either:

• Deployment configuration
• Region-specific backend issue

Regional Instability

You are correct to question the “instability” explanation after a week.

If it were purely transient, it would typically resolve within hours, not days.

Your test in swedencentral is a very good diagnostic step. If it works there, that confirms a region-specific issue in useast2.

Recommended Next Steps

Please verify

Confirm explicit deployment exists, Confirm you are using the deployment name, Confirm deployment SKU supports realtime, Test in swedencentral

There is currently no retirement notice for gpt-4o-transcribe. If it is listed as Generally Available in the portal, it should function once properly deployed and supported in-region.

Thank you!

Answer 2

Thanks for the reply — appreciate it.

If it is instability, I have experienced the issue for over a week now.
We are in useast2, and the gpt-4o-transcribe model is available according to the portal.

Do I need to explicitly deploy it?
The documentation doesn't allude to it, and I didn't with Whisper-1, which worked.
The documentation states that it is available generally.

I have deployed to swedencentral , and will test with it to see if it is a region specific issue.

Answer 3

Hi ,

Thanks for reaching out to Microsoft Q&A.

If whisper-1 works but gpt-4o-transcribe fails in useast2 with gpt-realtime, the issue is almost certainly regional model availability or backend instability, not your audio format.

gpt-4o-transcribe is newer than whisper-1 and is not consistently deployed across all Azure regions. In some regions (including useast2), it may be:

• Partially deployed

Behind a different capacity SKU

Experiencing backend instability

Not enabled for realtime transcription in that region

The server_error with no code strongly indicates a serviceside failure, not a payload issue. Your PCM 24kHz format is valid since it works with whisper-1.

Since it works against OpenAI-hosted realtime endpoints, that confirms:

Your websocket implementation is correct

Your audio encoding is correct

The failure is region/model specific in Azure

What to check:

Confirm gpt-4o-transcribe is officially available in useast2 for realtime scenarios (not just batch).

Try deploying in a different region (for example, eastus or swedencentral).

Verify your Azure OpenAI resource SKU supports realtime transcription with that model.

Check Azure Service Health for transient issues in useast2.

Please 'Upvote'(Thumbs-up) and 'Accept' as answer if the reply was helpful. This will be benefitting other community members who face the same issue.