Finetuning a model to "forget" all but some sentences

[joazoa]

Hello,

Thanks for the great project!
I have a bit of a weird question, I have a perfectly trained model but now need a model that works only for a few phrases and words and should forget everything else. The original model is trained on a lot of data and will handle many voices, accents and background noises.

The contextual biasing or the use of a language model would achieve that, but i want to dumb down the actual model so that it cannot be used for other sentences (because of commercial reasons).

If i just finetune on those few phrases, I am worried that because of the very limited data in the target trainset, the model will forget the variety of voices and accents.
What If i take the original dataset and model (trained with bpe) and I substitute all unwanted words with , would that work ? Would it work better if I retrain the original model with word based tokensfile, and then finetune on the same data with all unwanted words mapped to ?

Does anybody have any better or more creative ideas?