A model with 82 million parameters does the voice

It is roughly one four-thousandth the size of a frontier language model, and it sounds better than systems a hundred times larger. Speech turned out to be an easier problem than anyone assumed.