3 ms·

They're literally comparing it to OpenAI's open-sourced speech recognition model in the paper... Oh and of course their model is hidden behind an API.