In an avatar descriptor if you choose the viseme/blendshape option, click "auto detect" and do not provide every viseme, lip sync will not work.

This is not very helpful or intuitive. The SDK should at least have a message "you have not provided every viseme".