We propose a foundation model for unified speech generation with masked generative pre-training.