Will we? The resources to train these things, and to keep them up-to-date, seem like a pretty big moat that's likely to keep them owned by megacorps.
[EDIT] I don't (mostly) mean computing resources, but data. Including expensive, well-tagged sets.
[EDIT] I don't (mostly) mean computing resources, but data. Including expensive, well-tagged sets.
Local speech recognition has been a thing for decades, then megacorps started to do online speech recognition with expensive-to-train models which worked better.
Over time, both online and offline models have gotten better. But for most usecases, online models trained by megacorps still win.
Whisper and whisper.cpp have gotten us close to the tipping point.