I am looking for a speech-to-text model that is available for commercial use (such as Whisper and Chirp) of which full access to the model is possible and performs (near) SOTA. I would like to be able to access and modulate the distribution it predicts over the vocabulary per timestep that it predicts. Is this possible in Whisper or USM, and if not, what other options exist?
For example, I would like to develop a mechanism that can quantify certainty over a certain speech-to-text prediction on word level basis. Another thing I would like to develop is to apply a prior distribution to make a certain subset of the vocabulary more likely to occur a priori. That way I avoid having to finetune to my own data.
I'm using Pytorch.
Any help would be appreciated. Thanks in advance! :)
