MelGAN is a non-autoregressive, fully convolutional generative adversarial network (GAN) designed for conditional waveform synthesis. The paper introduces MelGAN, which effectively generates high-quality audio waveforms by using mel-spectrograms as input. Unlike previous methods that rely on autoregressive models or signal processing techniques, MelGAN achieves high-quality waveform generation without the need for perceptual loss functions or additional distillation techniques. It is significantly faster than existing models, with inference speeds over 100 times faster on a GPU and over 2 times faster on a CPU. The model is also capable of generalizing to unseen speakers for mel-spectrogram inversion. The MelGAN model consists of a generator and a discriminator. The generator is a fully convolutional network that upsamples mel-spectrograms to raw waveforms, while the discriminator is a multi-scale network that evaluates the quality of generated waveforms. The generator uses weight normalization for better training dynamics, and the discriminator uses grouped convolutions to allow larger kernel sizes with fewer parameters. The training objective combines hinge loss with a feature matching loss to ensure the generated waveforms are realistic and coherent. The model is evaluated on several tasks, including speech synthesis, music translation, and unconditional music synthesis. It achieves high-quality results in all cases, with the MelGAN model outperforming existing models in terms of both quality and speed. The model is also shown to generalize well to unseen speakers, demonstrating its effectiveness in learning speaker-invariant mappings from mel-spectrograms to raw waveforms. The paper also discusses the limitations of the model, including the requirement for time-aligned conditioning information and the challenges of feature matching with paired ground truth data. However, the model is shown to be effective in a wide range of tasks and is a promising approach for future research in audio synthesis.MelGAN is a non-autoregressive, fully convolutional generative adversarial network (GAN) designed for conditional waveform synthesis. The paper introduces MelGAN, which effectively generates high-quality audio waveforms by using mel-spectrograms as input. Unlike previous methods that rely on autoregressive models or signal processing techniques, MelGAN achieves high-quality waveform generation without the need for perceptual loss functions or additional distillation techniques. It is significantly faster than existing models, with inference speeds over 100 times faster on a GPU and over 2 times faster on a CPU. The model is also capable of generalizing to unseen speakers for mel-spectrogram inversion. The MelGAN model consists of a generator and a discriminator. The generator is a fully convolutional network that upsamples mel-spectrograms to raw waveforms, while the discriminator is a multi-scale network that evaluates the quality of generated waveforms. The generator uses weight normalization for better training dynamics, and the discriminator uses grouped convolutions to allow larger kernel sizes with fewer parameters. The training objective combines hinge loss with a feature matching loss to ensure the generated waveforms are realistic and coherent. The model is evaluated on several tasks, including speech synthesis, music translation, and unconditional music synthesis. It achieves high-quality results in all cases, with the MelGAN model outperforming existing models in terms of both quality and speed. The model is also shown to generalize well to unseen speakers, demonstrating its effectiveness in learning speaker-invariant mappings from mel-spectrograms to raw waveforms. The paper also discusses the limitations of the model, including the requirement for time-aligned conditioning information and the challenges of feature matching with paired ground truth data. However, the model is shown to be effective in a wide range of tasks and is a promising approach for future research in audio synthesis.