Research map / Image generation / GANs
BigVGAN: A Universal Neural Vocoder with Large-Scale Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the BigVGAN generator with periodic Snake activations and anti-aliased representation | HiFi-GAN-style vocoder · LibriTTS dev | PESQ: 2.947 -> 4.027 | Helped |
| Scale the generator from 14M to 112M parameters | BigVGAN · LibriTTS test | SMOS: 4.20±0.08 -> 4.26±0.08 | Helped |
| Use BigVGAN instead of HiFi-GAN | HiFi-GAN (V1) · Unseen languages, simulated noise | SMOS: 4.13±0.08 -> 4.26±0.07 | Helped |
| Use BigVGAN instead of HiFi-GAN | HiFi-GAN (V1) · MUSDB18-HQ | Average SMOS: 4.08±0.05 -> 4.26±0.04 | Helped |
| Add the anti-aliasing filter to Snake activations (ablation) | BigVGAN-base · MUSDB18-HQ | Average SMOS: 4.15±0.04 -> 4.20±0.04 | Helped |
| Add Snake activation (ablation) | BigVGAN-base without filter or Snake · LibriTTS dev-other | PESQ: 3.240 -> 3.364 | Helped |
| Add the anti-aliasing filter (ablation) | BigVGAN-base with Snake activation · LibriTTS dev-other | PESQ: 3.364 -> 3.464 | Helped |
| Train on diverse full LibriTTS data rather than train-clean-360 (ablation) | BigVGAN · LibriTTS dev-other | PESQ: 3.847 -> 4.001 | Helped |
| Train on diverse full LibriTTS data rather than VCTK (ablation) | BigVGAN · LibriTTS dev-other | PESQ: 3.818 -> 4.001 | Helped |
| Replace MSD with MRD (ablation) | HiFi-GAN-style generator · LibriTTS | audio quality | Helped |
| Replace transposed-convolution upsampling with anti-aliased upsampling (ablation) | BigVGAN · GAN training | training stability | Hurt |
| Use Snake activations in the discriminators (ablation) | BigVGAN · audio quality | perceptual quality | Hurt |
Closest papers by meaning
- High Fidelity Speech Synthesis with Adversarial Networks (ICLR 2020 · similarity 0.91)
- HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis (NeurIPS 2020 · similarity 0.90)
- Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis (ICLR 2024 · similarity 0.88)
- FINALLY: fast and universal speech enhancement with studio-like quality (NeurIPS 2024 · similarity 0.87)
- FlowDec: A flow-based full-band general audio codec with high perceptual quality (ICLR 2025 · similarity 0.87)
- DFlow: A Generative Model Combining Denoising AutoEncoder and Normalizing Flow for High Fidelity Waveform Generation (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.