BigVGAN

Few components are reused across modern speech stacks as often as BigVGAN, NVIDIA's universal neural vocoder that turns mel spectrograms into waveforms for speech, music, and environmental sound. Its generator introduces anti-aliased multi-periodicity composition with snake activations under log-scale parameterization, which is what lets a single model generalize to unseen speakers and recording conditions. The 2024 BigVGAN-v2 release added a custom CUDA inference kernel measured at 1.5 to 3x faster synthesis on an A100, discriminators using multi-scale sub-band CQT analysis, and pretrained checkpoints of 112M to 122M parameters covering 22 kHz to 44 kHz sampling rates and 80 to 128 mel bands, joining the earlier 14M parameter base models, all downloadable from Hugging Face. Running it takes PyTorch and Python 3.10, with a GPU strongly preferred for practical synthesis speed, and the repository includes training code and a local Gradio demo. Code and weights are MIT licensed, so commercial use is unrestricted, and the ICLR 2023 model remains a default vocoder choice cited throughout TTS and voice conversion research.