Translation to other languages; temperature, Top K for whisper (speech to text model)

[hoohooprog]

Hi

I've been wondering quite awhile now if it could be possible that translation to other language is possible because of the attention matrix, which may return the inverse of a language.

Thus if we wanted to translate from German -> french, it goes to English first using Attention, before finding French words for the resulting matrix?

I have not really experimented with such idea yet because my focus is on generating alternative phrases of the same meaning. So far beam search seems to do it literally but it seems temperature or top K might be the feature I'm looking for, which I think happens because of how the model is trained (masking). Just not quite so how it works for a speech to text model like whisper.