I mean, this is basically just a wrapper around whisper, which has been around for a while now, so IMO this is nothing revolutionary. And the “small” model used is actually quite inaccurate, but the large model is not capable of real-time use on a typical system.