NVIDIA has recently unveiled BigVGAN v2, a revolutionary generative AI model designed for zero-shot waveform audio generation, as reported by the NVIDIA Technical Blog. This cutting-edge model not only offers significant enhancements in speed and quality but also establishes itself as a state-of-the-art solution in the realm of audio generative AI.

The Breakthrough BigVGAN: A Game-Changer in Audio Generation

BigVGAN serves as a universal neural vocoder engineered to synthesize audio waveforms from Mel spectrograms. Featuring a fully convolutional architecture with multiple upsampling blocks and residual dilated convolution layers, the model boasts the anti-aliased multiperiodicity composition (AMP) module. This module is tailor-made for generating high-frequency and periodic sound waves, thereby minimizing artifacts in the synthesis process.

What’s New in BigVGAN v2?

BigVGAN v2 introduces a slew of improvements over its predecessor, including:

• Unprecedented audio quality across diverse metrics and audio types.
• Up to 3x faster synthesis speed facilitated by optimized CUDA kernels.
• Pretrained checkpoints for varied audio configurations.
• Support for a sampling rate up to 44 kHz, encompassing the highest audible human frequencies.

Empowering BigVGAN v2: Mapping Every Sound Imaginable

The advent of waveform audio generation presents significant opportunities in virtual worlds and has been a focal point of intense research. BigVGAN v2 overcomes earlier constraints, delivering top-notch audio with enhanced precision. Leveraging NVIDIA A100 Tensor Core GPUs and a dataset over 100 times larger than its forerunner, BigVGAN v2 is proficient in generating high-quality sound waves spanning speech, environmental sounds, and music.

Unlocking the Limits of Human Hearing: High-Frequency Sound Reproduction

Previous models were constrained to sampling rates between 22 kHz and 24 kHz, limiting their auditory range. Enter BigVGAN v2, which extends this range to 44 kHz, capturing the full human auditory spectrum. This breakthrough enables the model to recreate intricate soundscapes, from robust drums to crystalline cymbals in music.

Enhanced Efficiency: Accelerated Synthesis with Custom CUDA Kernels

BigVGAN v2 harnesses custom CUDA kernels for accelerated synthesis, achieving up to 3x faster inference than its predecessor. These kernels empower the generation of audio waveforms, delivering outputs up to 240 times faster than real-time on a single NVIDIA A100 GPU.

Delving into Audio Quality Metrics

BigVGAN v2 showcases superior audio quality for speech and general audio when compared to its predecessor. It also stands shoulder-to-shoulder with the Descript Audio Codec at a 44 kHz sampling rate, underscoring its prowess in producing exceptional waveforms across diverse audio categories.

In Conclusion

NVIDIA’s BigVGAN v2 heralds a new era in audio synthesis, setting new benchmarks in quality and efficiency across all audio types while covering the entire spectrum of human hearing. With synthesis speeds now up to 3x faster, this model is primed to cater to a multitude of audio configurations efficiently.

For in-depth insights, we invite you to explore the BigVGAN v2 model card on GitHub.

Image source: Shutterstock