NVIDIA/BigVGAN

Health v2 · maintenance only

Adoption not part of the score

1227 stars · 147 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

BigVGAN is NVIDIA's official PyTorch implementation of a universal neural vocoder (ICLR 2023) that generates high-fidelity raw audio waveforms from acoustic features using large-scale GAN training. It includes pretrained checkpoints up to 44 kHz, custom fused CUDA kernels for fast inference, and Hugging Face Hub integration.

Use cases

• convert mel spectrograms to audio waveforms with a neural vocoder
• synthesize high-quality speech audio for TTS pipelines
• generate singing voice and instrumental audio from acoustic features
• run zero-shot vocoding on unseen speakers, languages, and recording conditions
• accelerate vocoder inference with fused CUDA kernels on GPU
• fine-tune or train a GAN-based vocoder on custom audio datasets

When to choose

• you need a state-of-the-art universal vocoder that generalizes out-of-distribution without fine-tuning
• you want pretrained checkpoints supporting up to 44 kHz audio and 512x upsampling
• you need fast GPU inference with optimized CUDA kernels
• you want easy integration via Hugging Face Hub

When to avoid

• you need CPU-only inference at scale, since it is optimized for NVIDIA GPUs
• you need a full end-to-end text-to-speech system rather than a vocoder component
• you require non-PyTorch frameworks like TensorFlow or JAX

Facets

library · maturity stable

audio-processing machine-learning deep-learning tts speech-processing machine-learning deep-learning python cross-platform neural-vocoder gan speech-synthesis audio-generation pytorch cuda-kernel pretrained-models hugging-face audio gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="NVIDIA/BigVGAN")