Transcription Quality Issues with Fast-Transcription for German

Question

Has anyone observed a decline in transcription quality for non-English languages, specifically with Fast-Transcription?

Recently, it has been noted that Fast-Transcription tends to interpret spoken content as English, despite the de-DE locale being selected and speakers clearly not using English. This issue appears to be more prevalent than before.

Was there a recent model update, or are there settings available to ensure the transcription prioritizes the selected locale?

Answer 1

Hello Matthias !

Thank you for posting on Microsoft Learn.

I have seen many people reporting similar issues recently with Azure AI Speech Fast-Transcription, particularly for non-English locales like in your case.

Microsoft may have rolled out an update to the Fast-Transcription models. Some updates can unintentionally shift the model’s behavior (fallback to English bias), especially if multilingual handling or language auto-detection features were modified.

Even when specifying a locale like de-DE, Fast-Transcription may still perform internal language confidence checks. If audio quality, accents, or noise cause confusion, it may fallback to interpreting content as English especially if the model was optimized recently for English-first use cases.

If you're using auto-detection or multi-locale settings, the transcription engine may incorrectly classify the speech language. Even though de-DE is selected, it may not be strictly enforced unless configured properly.

You can try to explicit the Set Locale in API Call and that it is not relying on auto-detection :

{
  "locale": "de-DE"
}

Because when you force the locale strictly (using the Speech SDK with SpeechRecognizer(recognizerConfig, "de-DE")) can reduce misclassification.

If you're using AutoDetectSourceLanguageConfig, try switching to a static locale config since auto-detect can misclassify speech as English even when the correct locale is available.

You may need to contact support for more help.