open-mmlab/Amphion

Amphion implements task-specific generation pipelines (TTS, SVC, voice conversion, text-to-audio) with unified architectural components including neural vocoders and standardized evaluation metrics for reproducibility. Built on PyTorch, it provides modular design supporting both classical and foundation models—such as Vevo (zero-shot voice imitation with prosody control) and MaskGCT (non-autoregressive TTS)—alongside large-scale datasets like Emilia (200k+ hours) for training. The toolkit integrates with Hugging Face and ModelScope, enabling seamless model sharing and deployment across speech generation tasks.

9,712 stars. No commits in the last 6 months.