descriptinc/melgan-neurips

Implements a non-autoregressive, fully convolutional architecture with architectural innovations and training techniques that enable stable GAN-based waveform generation, achieving 100x+ faster-than-realtime synthesis on GPUs. Generalizes to unseen speakers for mel-spectrogram inversion and demonstrates effectiveness across speech synthesis, music domain translation, and unconditional generation tasks. Available via PyTorch Hub with minimal dependencies for integration into end-to-end TTS pipelines.

1,037 stars. No commits in the last 6 months.