[sbrother]

right, but either whatever audio tokenization it's doing doesn't seem to encode pitch, or there's ~nothing where pitch is relevant in the training set.