Translation for account number bringing back English mix

[tsmith-ng]

I have audio of users giving me thier account number which is a mix of letters and numbers. For the most part, audio.translation does fine when they are spoken in English. But when I try audio in French, it is all over the place. Funny that translation API doesn't take language!

Example: 12BB99 (une deux "bae" "bae" neuf neuf)
Returned: 12baby99

What am I doing wrong? Thanks!

[glangford]

whisper will try to identify the language based on the first 30s of audio (if you don't specify the language). But it isn't always correct.

If you know the language of the audio is French, try specifying language='fr'. If you don't know the language, the answer here might be helpful:

• How do i fine tune whisper to transcribe the numbers properly #1982

You could also try using initial_prompt and/or a larger model.
https://cookbook.openai.com/examples/whisper_prompting_guide

[tsmith-ng]

Thanks! For the translation API there isn't a language property. You can only set that on the transcribe method. I will take a look at the initial_prompt though.

[tsmith-ng]

What is also strange is that it properly translates the 1 2 and then messes up the letters and then gets the 9 9 right.

[tsmith-ng]

It could be hearing the BB as "bébé" which is French for "baby"