Amphion

Hosted under the OpenMMLab organization on GitHub, Amphion collects reference implementations of many speech and audio generation models in one PyTorch codebase for reproducible comparison. Recipes cover text-to-speech (FastSpeech2, VITS, VALL-E, NaturalSpeech2, MaskGCT), zero-shot voice and accent conversion through the Vevo family, singing voice synthesis and conversion, latent-diffusion text-to-audio, neural codecs like DualCodec, and GAN, flow, diffusion and autoregressive vocoders. It also bundles objective metrics such as F0 error, speaker similarity and FAD, dataset preprocessing, the Emilia pipeline for turning in-the-wild recordings into training data, and SingVisio, an interactive view into a diffusion-based singing converter. Setup is a conda environment plus a shell script, or a CUDA Docker image; training needs a capable NVIDIA GPU, while Hugging Face checkpoints allow inference-only tests. The code is MIT and usable commercially, but several released weights, including MaskGCT and Vevo, are CC BY-NC, so read each model card. Best for researchers who want comparable baselines rather than a finished app.