Is it possible to prompt the model with instructions, like Unmute?
#1
by deathknight0 - opened
Thank you for your release. Just wondering if this repo supports prompting the model natively - like what you implemented in Unmute.
Hi, what would that mean for a speech-to-text? It is possible to feed in a prefix of audio + transcript which would condition the model to transcribe words in certain ways. For that you'd want to hack around in the Pytorch implementation, see here, we don't provide a high-level interface for that atm. But if you mean something like "transcribe every proper name in upper case", that is not possible.