MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis

TL;DR

MelGAN is a non-autoregressive, fully convolutional GAN model that generates high-quality, coherent audio waveforms efficiently for tasks like speech synthesis and music, outperforming previous models in speed and quality.

Contribution

The paper introduces MelGAN, a novel GAN architecture with architectural innovations and training techniques that enable fast, high-quality waveform synthesis across multiple audio domains.

Findings

Achieves high MOS scores in mel-spectrogram inversion

Runs over 100x faster than real-time on GPU

Generalizes well to unseen speakers and domains

Abstract

Previous works (Donahue et al., 2018a; Engel et al., 2019a) have found that generating coherent raw audio waveforms with GANs is challenging. In this paper, we show that it is possible to train GANs reliably to generate high quality coherent waveforms by introducing a set of architectural changes and simple training techniques. Subjective evaluation metric (Mean Opinion Score, or MOS) shows the effectiveness of the proposed approach for high quality mel-spectrogram inversion. To establish the generality of the proposed techniques, we show qualitative results of our model in speech synthesis, music domain translation and unconditional music synthesis. We evaluate the various components of the model through ablation studies and suggest a set of guidelines to design general purpose discriminators and generators for conditional sequence synthesis tasks. Our model is non-autoregressive, fully convolutional, with significantly fewer parameters than competing models and generalizes to unseen speakers for mel-spectrogram inversion. Our pytorch implementation runs at more than 100x faster than realtime on GTX 1080Ti GPU and more than 2x faster than real-time on CPU, without any hardware specific optimization tricks.