BigVGAN

BigVGAN is a universal neural vocoder: a Generative Adversarial Network-based model that turns Mel Spectrogram into raw audio waveforms, and is trained at large scale so it generalises to speakers, languages and recording conditions it didn't see in training.

It builds on HiFi-GAN, and replaces the usual activations with the Snake Activation Function, which adds a Periodic Inductive Biases that suits audio.

Idea from paper BigVGAN: A Universal Neural Vocoder with Large-Scale Training.