Why is Whisper transcribing my podcast in Welsh and not returning extra file types?

Below are some points to consider regarding your observations:

Welsh-like Output

• Language Detection & Noisy Audio:
  Whisper is designed to automatically detect language unless you explicitly specify one. In cases where the audio is very noisy or has atypical characteristics (like wind noise), the language-detection component may go awry and mistakenly output text that looks like another language. The “Welsh” output you saw is likely an artifact of misclassification rather than an intentional translation.

• Transcript vs. Model Input:
  As detailed in the documentation for realtime transcription, the model that handles audio natively (for internal processing) can diverge somewhat from the separate ASR (Whisper) transcription output. This means that even if the model understands the audio correctly, the transcription you see might not perfectly align with what the model processed because of how Whisper interprets the input.

Changes in Output Files

• Current Output Format:
  Previously, Whisper might have produced multiple file types (e.g., subtitles with timing or word-level timestamps) in addition to the plain transcription. However, recent versions of the API now return a JSON payload with either:

  • A standard transcription object, or
  • A verbose transcription object when requested.

• Discontinuation of Extra Files:
  The additional outputs like separate subtitle files or detailed time-series are no longer provided directly in the API response JSON. If needed, some of these details (such as word or segment-level timestamps) may be enabled using specific options (depending on how you set up your API request) or might require post-processing of the verbose output.

What to Do Next

• Explicit Language Parameter:
  When making your API requests, set the language parameter explicitly if you are always expecting English. This may help avoid misclassification in cases where the audio is less clear.

• Handling Verbose Output:
  If you require additional timing details or subtitles, check the current API documentation to see which options are offered in the verbose transcription objects. Adjust your request options accordingly.

---

If you’re still encountering inconsistent behavior or need more granular output, it might be worth consulting the latest API guidelines or reaching out to OpenAI support since changes in API behavior can sometimes take users by surprise.