this is more question than an issue

[silvacarl2]

hi, we have 12M names and we would like to fine tune whisper on them. also, i am happy to share with you the results.

the question is it better to fine tune whisper using the entire spoken name? Or is it better to fine tune using invidial names and recording snippets of each anme spoken?

[sanchit-gandhi]

Hey @silvacarl2! Sorry for the late reply here! The best option would be to fine-tune on the closest scenario to what you expect at inference time. If you expect the model to transcribe the entire spoken name, then you should go with that.

[silvacarl2]

excellent. we now have 14M samples to use for fine tuning. yes i think entire spoken name is the way to go.

Any pointers to the latest best techniques to fine tuning whisper for a use case like this?