Getting speaker: '@' from the gpt-40-transcribe-diarize model

Surprised I couldn’t find anyone else that has ran into this bug.

I’m getting, speaker assignment as ‘@’:

{
“end”: 93.026,

“speaker”: “@”,

“start”: 89.326,

“text”: “Yes, you can. Yes, you can. Yes, you can.”

    }

Key Observations:

  • The text was not present in the actual audio file, it repeated the same chunk of text mutiple times
  • The start and end of think chunk specifically was after the end of the actual audio (audio file was 89 secs, this chunk ends at 93)
  • Ran the same audio file through the gpt-4o-transcribe-diarize model and consistenly got the speaker: ‘@’

Wasen’t able to find anything related to this issue in the docs.