kan-bayashi/ParallelWaveGAN

Implements five non-autoregressive vocoder architectures (Parallel WaveGAN, MelGAN, Multiband-MelGAN, HiFi-GAN, StyleMelGAN) optimized for real-time mel-spectrogram-to-waveform synthesis with modular generator-discriminator combinations and STFT loss functions. Designed for seamless integration with ESPnet-TTS and Tacotron2-based systems, supporting both text-to-speech and singing voice synthesis through pre-built recipes for diverse datasets (LJSpeech, VCTK, LibriTTS, Kiritan, Opencpop, etc.). Provides distributed multi-GPU training via PyTorch with optional NVIDIA Apex support and includes inference utilities for deployment.

1,637 stars. No commits in the last 6 months.