I have been craving for a small, extremely accurate voice model for a year. Having that would be an absolute game changer because I think that's the inflection point where people start interacting with their devices primarily through audio.
If devices were truly private, we could simply set our mics to always-on and massive Personal Wikis would build themselves over time. Imagine that, an autobiography for everyone. Of course the downside is massive, arguably useless storage for humans. Would be pretty valuable training data though.