MelGAN

MelGAN is a non-autoregressive, fully convolutional Generative Adversarial Network-based vocoder that converts Mel Spectrogram into raw audio waveforms.

The generator is trained against a multi-scale waveform discriminator, which judges the audio at several different resolutions. Because it isn't autoregressive, it's much faster than models like WaveNet: A Generative Model for Raw Audio, and later vocoders like HiFi-GAN build on the same approach.

Model from paper MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis.