I mean, it depends on your hardware and the model’s size/intelligence.

Worst case for me is many seconds of preprocessing followed by 4-5 words a second.

But you can get almost instant responses + way faster than you can read too.