I think adaptation isn't working for Speech API v2 at all. I am using the "long" model. It doesn't throw the same exception you're seeing but the phrase sets are ignored. The same code works in v1. This is unfortunate.