LLMs are so good at cleaning up speech to text. If you have not tried this application you should. I think it’s the one place where this dubious tech really feels “revolutionary” —But, to mitigate my guilt over the cavalier way most cloud models are trained I want to embark on a journey of training my own model for this purpose: curate the training data, run it locally. Then I want to enjoy dictation that can pick up it is “ant” not “aunt” from context. “trophallaxis” not “trophies lacking” etc.