Fast transcription with diarization times out (HTTP 408 InternalCommunicationFailed) in Italy North since Oct 13, 2026

Since 2026-10-13 around 08:00 UTC, every fast transcription request with diarization enabled fails in the Italy North region. The service waits about 60 seconds and then returns HTTP 408 with the body {"code":"InternalCommunicationFailed","message":"The request has timed out."}. With api-version 2024-05-15-preview and the preview diarizationSettings schema, the same request returns HTTP 500 after about 63 seconds.

Request: POST https://<resource>.cognitiveservices.azure.com/speechtotext/transcriptions:transcribe?api-version=2025-10-15 as multipart/form-data, with audio = katiesteve.wav and definition = {"locales":["en-US"],"diarization":{"enabled":true,"maxSpeakers":2}}

What I verified: (1) It fails with any audio, including a 4-second clip and the Microsoft sample katiesteve.wav from the Speech SDK repo. (2) It fails with api-version 2024-11-15 and 2025-10-15, with maxSpeakers set to 2 or 3 or omitted, and with or without locales. (3) The same request without diarization succeeds in a few seconds on the same resource. (4) It fails on two different resources in Italy North: an existing AI Services resource (S0) and a brand new Speech resource (F0) created only to reproduce the issue, so it is not specific to one resource. (5) The same request with diarization succeeds on Sweden Central (2 speakers detected in about 5 seconds). (6) Batch transcription with diarization works in Italy North. (7) Nothing changed on our side: same code, same configuration, same Entra ID authentication; the last successful request was on 2026-10-12 at 19:06 UTC. (8) Resource Health shows no explanation and the Azure status page shows no incident.

Impact: we run a production pipeline that needs speaker diarization and has to stay in Italy North for data residency reasons. As a temporary workaround we moved fast transcription to Sweden Central, and we would like to move it back as soon as possible.

Question: is there a known issue with the diarization backend of fast transcription in Italy North? Could someone from the Speech team look into it? I can provide request IDs and timestamps privately.