The speech is so fake sounding. Not fake technically, but like a fake/pacifying kind of person. It is a very strange tone to train for.

The model is not interested in the conversation, it is just "serving" a conversation.

Which is very different from the engagement of SOTA text models.