Fast Transcription returns English words for short Polish speech under forced pl-PL (regression since Aug 3, 2026)

We use the Azure Speech Fast Transcription REST API with a single, forced locale "locales": ["pl-PL"]:

POST https://westeurope.api.cognitive.microsoft.com/speechtotext/transcriptions:transcribe?api-version=2024-11-15

Since August 3, 2026, short Polish words are transcribed as unrelated English words. For example, the Polish word "nie" (meaning "no") comes back as the English word "Yeah" (meaning "yes") — both the language and the meaning are wrong. Longer Polish utterances transcribe correctly; only short, acoustically ambiguous phrases are affected.

This is not a language identification issue. The response phrase has the correct locale: "pl-PL", yet the recognized text is English, with very low confidence:

json:
{ "offsetMilliseconds": 430, "durationMilliseconds": 400, "text": "Yeah.", "words": [ { "text": "Yeah.", "offsetMilliseconds": 430, "durationMilliseconds": 400 } ], "locale": "pl-PL", "confidence": 0.20116459 }

Note the pinned api-version=2024-11-15 (unchanged for a long time) and locale: "pl-PL" — so this is not caused by an api-version change or by language identification on our side. It worked correctly until August 3, which points to a service-side model change under the same API version.

Questions:

1. Is there a known change to the Fast Transcription model rolled out around August 3, 2026 affecting short-utterance recognition for pl-PL in West Europe?
2. How do we get correct recognition for short Polish words under a forced pl-PL locale?
3. If this is a regression, what is the right path to escalate it to the product team?

I can privately share the source audio and full request/response if an engineer picks this up.