Sub-Billion Parameter Models

Recognize that many current audio generative models use less than a billion parameters, making them easier to train and customize locally.