Research map / Image generation / Autoregressive and latent generation
High-Fidelity Audio Compression with Improved RVQGAN
Rithesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, Kundan Kumar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| replace ReLU with Snake activations (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | SI-SDR: 6.92 -> 9.12 | Improved |
| remove the multi-band STFT discriminator (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | objective metrics: 1.82 -> 1.80 STFT distance | No clear effect |
| remove adversarial losses (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | SI-SDR: 9.12 -> 1.07 | Hurt |
| remove adversarial losses (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | bitrate efficiency: 99% -> 62% | Hurt |
| replace multi-period discriminator with single-scale waveform discriminator (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | SI-SDR: 9.12 -> 8.51 | Hurt |
| replace multi-scale mel loss with single-scale high-hop mel reconstruction (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | SI-SDR: 9.12 -> 7.68 | Hurt |
| use a codebook latent dimension of 8 instead of 2 or 256 (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | SI-SDR: 2.22 -> 9.12; 5.09 -> 9.12 | Improved |
| replace projected lookup with EMA codebook learning (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | SI-SDR: 9.12 -> 8.33 | Hurt |
| apply quantizer dropout rather than no dropout (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps and lower bitrates | reconstruction quality across bitrates: 10.14 -> 9.74 SI-SDR | Mixed |
| increase maximum bitrate from 8 kbps to 24 kbps (ablation) | Improved RVQGAN · 44.1 kHz audio | SI-SDR: 10.75 -> 13.83 | Improved |
| remove balanced data sampling (ablation) | Improved RVQGAN · 44.1 kHz audio, 8 kbps | STFT distance: 1.82 -> 1.94 | Hurt |
| use Improved RVQGAN rather than EnCodec at the tested 24 kbps configuration | 24 kHz audio compression · 24 kHz audio, 24 kbps | SI-SDR: 9.69 -> 16.40 | Improved |
Closest papers by meaning
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.87)
- ALMTokenizer: A Low-bitrate and Semantic-rich Audio Codec Tokenizer for Audio Language Modeling (ICML 2025 · similarity 0.87)
- MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound Generation (ICLR 2025 · similarity 0.86)
- Compression via Pre-trained Transformers: A Study on Byte-Level Multimodal Data (ICML 2025 · similarity 0.86)
- DFlow: A Generative Model Combining Denoising AutoEncoder and Normalizing Flow for High Fidelity Waveform Generation (ICML 2024 · similarity 0.86)
- Scaling Transformers for Low-Bitrate High-Quality Speech Coding (ICLR 2025 · 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.
