Wow, Smalltalk. Smalltalk running on little tin. That is not the environment I would expect any kind of signal processing application to grow out of. Sound programs have the strangest genealogies.
Capytalk (the language underpinning Kyma) is like the basque of sound design languages. The mother tongue of most competitors is Max, which begat MaxMSP and PureData. I never spent much time in Max because it won't do real-time processing, which means it's un-musical. At the time, NI Reaktor was proprietary, teutonic and mostly useful for modeling analog synthesis but in the intervening years, Kyma has languished while Reaktor has flourished. I'll probably go that way. Although if someone could figure out a way to make PureData run hardware-accelerated on non-proprietary shit I'd be all over it.
I played with Pure Data a bit working through Miller Puckette's book. I'm not intimately familiar with it and this is far from my area, but I think it being CPU-bound has more to do with them not caring than it being hard. Jack will talk to anything CoreAudio (or ALSA on linux, or whatever on Windows) will talk to, and SuperCollider uses it too.