Solutions have improved from a couple years ago, but in general, the issue I've found is that azimuth encoding wasn't working too well.

Maybe by interpolating on a lot of dynamically recorded impulse responses, that can be done better.