An all-in-one generative Al model for speech