An Arabic voice sent to whisper-1 results in Hebrew text output.
For an Arabic speaker to get a Hebrew response is unacceptable. This is a major issue given the current middle east tension.
curl --location 'https://api.openai.com/v1/audio/transcriptions' \
--header 'Authorization: Bearer {your token}' \
--form 'file=@"/audio-in-AR.ogg"' \
--form 'model="whisper-1"'
In can be consistently reproduced with attached audio file.
In a way to propose a mitigation/fix against this, the Whisper API should allow a header to accept multiple “expected” languages. For example: expected-languages=en,ar
This way, Whisper would throw an error if the transcribed text contains characters other than those of the en,ar vocabulary.
Can anyone help this post to reach the whisper-api team?