NVIDIA has announced the release of BigVGAN v2, a groundbreaking generative AI model for zero-shot waveform audio generation, according to the NVIDIA Technical Blog. The new model delivers significant improvements in speed and quality, positioning itself as a state-of-the-art solution in the field of audio generative AI.

BigVGAN: A Universal Neural Vocoder

BigVGAN is a universal neural vocoder designed to synthesize audio waveforms from Mel spectrograms. The model employs a fully convolutional architecture with several upsampling blocks and residual dilated convolution layers. A key feature is the anti-aliased multiperiodicity composition (AMP) module, which is optimized for generating high-frequency and periodic sound waves, reducing artifacts in the process.

Improvements in BigVGAN v2

BigVGAN v2 introduces several enhancements over its predecessor:

• State-of-the-art audio quality across various metrics and audio types.
• Up to 3x faster synthesis speed through optimized CUDA kernels.
• Pretrained checkpoints for diverse audio configurations.
• Support for a sampling rate up to 44 kHz, covering the highest frequencies audible to humans.

Generating Every Sound in the World

Waveform audio generation is crucial for virtual worlds and has been a significant focus of research. BigVGAN v2 addresses previous limitations by delivering high-quality audio with enhanced fine details. Trained using NVIDIA A100 Tensor Core GPUs and a dataset over 100 times larger than its predecessor, BigVGAN v2 can generate high-quality sound waves from various domains, including speech, environmental sounds, and music.

Reaching the Highest Frequency Sound the Human Ear Can Detect

Previous models were limited to sampling rates between 22 kHz and 24 kHz. BigVGAN v2 extends this range to 44 kHz, capturing the entire human auditory spectrum. This allows the model to reproduce comprehensive soundscapes, from robust drums to crisp cymbals in music.

Faster Synthesis with Custom CUDA Kernels

BigVGAN v2 also features accelerated synthesis speed, using custom CUDA kernels to achieve up to 3x faster inference than the original BigVGAN. These kernels enable the generation of audio waveforms up to 240 times faster than real-time on a single NVIDIA A100 GPU.

Audio Quality Results

BigVGAN v2 shows superior audio quality for speech and general audio compared to its predecessor, as well as comparable results to the Descript Audio Codec at a 44 kHz sampling rate. This demonstrates the model’s capability to produce high-quality waveforms across various audio types.

Conclusion

NVIDIA’s BigVGAN v2 sets a new benchmark in audio synthesis, achieving state-of-the-art quality across all audio types and covering the full range of human hearing. The model’s synthesis speed is now up to 3x faster, making it highly efficient for diverse audio configurations.

For more information, users are encouraged to review the BigVGAN v2 model card on GitHub.

Image source: Shutterstock