Research map / Image generation / Autoregressive and latent generation
From Discrete Tokens to High-Fidelity Audio Using Multi-Band Diffusion
Robin San Roman, Yossi Adi, Antoine Deleforge, Romain Serizel, Gabriel Synnaeve, Alexandre Défossez
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 |
|---|---|---|---|
| remove the frequency EQ processor (ablation) | MBD · reconstruction 6kbps | ViSQOL: 3.38 -> 3.67 | Improved |
| use linear noise schedule instead of power schedule (ablation) | MBD · reconstruction 6kbps | ViSQOL: 3.67 -> 2.93 | Hurt |
| use cosine noise schedule instead of power schedule (ablation) | MBD · reconstruction 6kbps | ViSQOL: 3.67 -> 3.29 | Hurt |
| use a single full-band model instead of four band-specific models (ablation) | MBD · reconstruction 6kbps | ViSQOL: 3.67 -> 3.32 | Hurt |
| remove EQ processor, effect on high-frequency Mel-SNR (ablation) | MBD · reconstruction 6kbps | Mel-SNR-H: 8.46 -> 9.26 | Improved |
| increase sampling steps from 10 to 20 (ablation) | MBD · reconstruction 6kbps | ViSQOL: 3.61 -> 3.67 | Improved |
| increase sampling steps from 20 to 500 (ablation) | MBD · reconstruction 6kbps | ViSQOL: 3.67 -> 3.70 | No clear effect |
| increase sampling steps from 20 to 100 (ablation) | MBD · reconstruction 6kbps | ViSQOL: 3.67 -> 3.64 | No clear effect |
| replace EnCodec decoder with MBD (ablation) | EnCodec 6kbps · speech MUSHRA | MUSHRA: 79.03 -> 84.68 | Improved |
| replace EnCodec decoder with MBD | EnCodec 6kbps · music MUSHRA | MUSHRA: 84.67 -> 83.61 | No clear effect |
| replace EnCodec decoder with MBD | EnCodec 1.5kbps · speech MUSHRA | MUSHRA: 49.51 -> 65.83 | Improved |
| replace EnCodec decoder with MBD | EnCodec 3kbps · music MUSHRA | MUSHRA: 85.55 -> 87.31 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PriorGrad — PriorGrad: Improving Conditional Denoising Diffusion Models with Data-Dependent Adaptive Prior (ICLR 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Fast Timing-Conditioned Latent Audio Diffusion (ICML 2024 · similarity 0.88)
- MMDisCo: Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation (ICLR 2025 · similarity 0.87)
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.87)
- A Versatile Diffusion Transformer with Mixture of Noise Levels for Audiovisual Generation (NeurIPS 2024 · similarity 0.87)
- Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction (ICML 2025 · similarity 0.86)
- MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound Generation (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.
