Research map / Image generation / Autoregressive and latent generation
DiffWave: A Versatile Diffusion Model for Audio Synthesis
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 DiffWave LARGE instead of WaveNet | Neural vocoding on LJ Speech · LJ Speech neural vocoding | MOS: 4.43 -> 4.44 | No clear effect |
| Reduce DiffWave BASE inference to 6 fast-sampling steps (ablation) | DiffWave BASE, T=50 · LJ Speech neural vocoding | MOS: 4.38 -> 4.37 | No clear effect |
| Reduce DiffWave LARGE inference to 6 fast-sampling steps (ablation) | DiffWave LARGE, T=200 · LJ Speech neural vocoding | MOS: 4.44 -> 4.42 | No clear effect |
| Increase BASE diffusion steps from 20 to 50 (ablation) | DiffWave BASE · LJ Speech neural vocoding | MOS: 4.31 -> 4.38 | Helped |
| Use DiffWave instead of WaveGAN | Unconditional waveform generation · SC09 unconditional generation | MOS: 2.03 -> 3.39 | Helped |
| Use DiffWave instead of WaveGAN | Unconditional waveform generation · SC09 unconditional generation | FID: 1.349 -> 1.287 | Helped |
| Use DiffWave instead of WaveNet-256 | Unconditional waveform generation · SC09 unconditional generation | mIS: 10.0 -> 59.4 | Helped |
| Condition generation on digit labels | DiffWave · SC09 generation | MOS: 3.39 -> 3.50 | Helped |
| Use DiffWave instead of WaveNet-256 | Class-conditional generation · SC09 class-conditional generation | Accuracy: 60.70% -> 91.20% | Helped |
| Make DiffWave deeper and thinner: 48 layers and 128 channels (ablation) | DiffWave · SC09 class-conditional generation | Accuracy: 91.20% -> 94.00% | Helped |
| Make DiffWave deeper and thinner: 48 layers and 128 channels (ablation) | DiffWave · SC09 class-conditional generation | MOS: 3.50 -> 3.44 | Hurt |
| Increase WaveNet dilation cycle and depth (ablation) | WaveNet unconditional model · SC09 unconditional generation | audio quality | Hurt |
Closest papers by meaning
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.92)
- DFlow: A Generative Model Combining Denoising AutoEncoder and Normalizing Flow for High Fidelity Waveform Generation (ICML 2024 · similarity 0.88)
- PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation (ICLR 2025 · similarity 0.88)
- DMOSpeech: Direct Metric Optimization via Distilled Diffusion Model in Zero-Shot Speech Synthesis (ICML 2025 · similarity 0.87)
- Fast Timing-Conditioned Latent Audio Diffusion (ICML 2024 · similarity 0.87)
- MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound Generation (ICLR 2025 · similarity 0.87)
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.