Prior to the release of GPT4o (don't really know if it is related or not), Whisper API was able to translate properly text & musics, which is very cool... Since almost the new releases from OpenAI, it became totally incompetent.
For example try to send it the title "Paradise" from Coldplay and you will obtain this translation : "My Outro For My 20th Birthday. Thank you for watching! Please subscribe and Thumbs Up!"
I have tried to force the language to "en" because the song start after the first 30s and I obtained this : "You're still here?
...
...
...
...
...
...
..."
Finally I tried with an HD version of the file and I get musical notes as glyphs instead of text.
It has to be mentionned that it was working pretty well a few weeks ago with the same file.
As a proof, you can send the file to a whisper model available on replicate, setting language to english and you'll obtain the beautiful transcription "When she was just a girl, she expected the world. But it flew away from her reach, so she ran away in her sleep. Dreamed of para, para, paradise...."
#Whisper api has become incompetent since the release of gpt 4o
vague nymph
blazing owlBOT
@vague nymph