Actively working on it. I've not noticed any performance problems even with the large model (though the plan was always to run the speech recognition on a GPU

Actively working on it. I've not noticed any performance problems even with the large model (though the plan was always to run the speech recognition on a GPU - your use case may differ). It seems to be doing fairly well even with slightly noisy inputs, and certainly has better bang/$ than other non-API solutions that service my native language. While true real-time would definitely be nice, I can approximate it well enough with various audio slicing techniques.