MaskGCT

MaskGCT removes the two crutches most zero-shot TTS systems lean on, explicit text-to-speech alignment and phone-level duration prediction, by adopting a mask-and-predict paradigm over discrete tokens. Stage one predicts semantic tokens from text, stage two turns semantic tokens into acoustic tokens with a masked generative codec transformer, and parallel decoding fills in masked positions over a handful of steps rather than one token at a time. Trained on the 100,000 hour Emilia corpus of in-the-wild English and Chinese speech, it delivers zero-shot voice cloning with controllable duration and reached state-of-the-art similarity and quality results, earning an ICLR 2025 acceptance. The model ships inside the MIT-licensed Amphion toolkit from open-mmlab: setup is a repository clone, espeak-ng, and a pip requirements install, with checkpoints for the semantic codec, acoustic codec, T2S, and S2A stages on Hugging Face plus a Gradio demo. The released weights are CC-BY-NC-4.0, so commercial use is off the table, but for research it is one of the most complete open non-autoregressive TTS stacks.