MAI-Transcribe-2 with diarization returns 408 RequestTimeout even for 10-second audio (Southeast Asia)

No public Azure incident currently confirms a Southeast Asia MAI-Transcribe-2 degradation or provides an ETA, so check the subscription-scoped Service Health event, open a production-impacting support request, and fail over persistent requests to a separate Speech resource in another MAI-supported region.

Incident status and ETA

The public Azure Status feed currently shows no active event specifically covering MAI-Transcribe-2, Southeast Asia, or these 408 timeouts.^1^ That does not exclude a regional, subscription-specific, or emerging issue: Azure Status primarily shows incidents with broad impact, while Azure Service Health provides events relevant to your subscriptions and regions.^2^

Check:

1. Azure portal → Service Health → Service issues.
2. Filter for the affected subscription, Azure AI Services/Speech, and Southeast Asia.
3. Also check Health history, because resolved events are retained there.
4. Open any matching event to obtain its tracking ID, latest update, and estimated mitigation time if Microsoft has published one.^3^

There is no independently documented ETA available here. Any reliable ETA would need to come from the Service Health event or Azure Support.

Your results strongly warrant service-side investigation: standard transcription remains fast, while the MAI path is severely delayed and diarization consistently reaches the approximately 240-second timeout even for ten seconds of audio. This also distinguishes the case from the previously reported longer-recording diarization behavior.^4^

Immediate failover

MAI transcription is documented in these regions:

• centralindia
• eastus
• northeurope
• southeastasia
• westus
• westus2^5^

For an Asia-based workload, centralindia is the other documented MAI region in the same broad geography, but select the secondary according to your latency, compliance, and data-residency requirements. Test it before shifting production traffic.

A region change requires a new Foundry resource for Speech in the secondary region. Speech keys are region-scoped, so the Southeast Asia key must not be sent to another region’s endpoint.^5^

1. Create a Foundry resource for Speech in a supported secondary region.
2. Obtain that resource’s endpoint and key.
3. Submit the same ten-second test to the secondary endpoint using API version 2025-10-15.
4. Confirm both diarization and expected latency.
5. Redirect queued meeting requests only after the probe succeeds.

curl --location \
  'https://<secondary-resource>.cognitiveservices.azure.com/speechtotext/transcriptions:transcribe?api-version=2025-10-15' \
  --header 'Ocp-Apim-Subscription-Key: <secondary-key>' \
  --form 'audio=@"repro.wav"' \
  --form 'definition={
    "enhancedMode":{"enabled":true,"model":"MAI-Transcribe-2"},
    "diarization":{"enabled":true}
  }'

The API is synchronous, uses multipart form data, and returns 200 OK on success.^6^ Microsoft’s Speech resiliency guidance recommends retrying an isolated connectivity failure, but redirecting persistent timeout or service-unavailability failures to another region where a separate Speech resource is already provisioned.^7^

Avoid repeatedly retrying each request for another four minutes during this incident. Use a low-frequency health probe and open the circuit to Southeast Asia after persistent failures.

Escalation details

Because transcription has been blocked organization-wide for hours, create a technical support request from the affected Speech resource:

Speech resource → Support + Troubleshooting → Create a support request.^8^

Include:

• Resource ID and region, with secrets removed.
• UTC onset and degradation timeline.
• API version and endpoint operation.
• Exact HTTP status and response body.
• Request/correlation IDs and response headers from representative failures.
• The three-way comparison: standard, MAI without diarization, and MAI with diarization.
• Secondary-region test results.
• Business impact and current backlog.
• A minimal WAV, MP3, or FLAC reproduction file if possible, because those are the formats explicitly listed for MAI-Transcribe; this prevents OGG support from obscuring the timeout investigation.^9^

Creating a support request requires Owner, Contributor, Support Request Contributor, or equivalent Microsoft.Support/* permissions at subscription scope.^8^

References

1. Azure status
2. Azure status overview
3. Azure Service Health portal
4. MAI-Transcribe-2: speaker diarization fails on recordings over ~15 ...
5. Supported regions for Azure Speech
6. Transcriptions - Transcribe
7. Back up and recover speech customer resources
8. Create an Azure support request
9. MAI-Transcribe in Azure Speech (preview) (programming-language-csharp)