Research map / Image generation / Autoregressive and latent generation
DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism
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 |
|---|---|---|---|
| replace Gaussian-noise initialization with shallow diffusion from the auxiliary decoder prediction (ablation) | DiffSinger · PopCS test | MOS: 0.14 MOS gains | Helped |
| replace Gaussian-noise initialization with shallow diffusion from the auxiliary decoder prediction (ablation) | DiffSinger · PopCS inference | RTF: 45.1% speedup (RTF 0.191 vs. 0.348) | Helped |
| add shallow diffusion to the naive diffusion model (ablation) | DiffSinger · PopCS CMOS evaluation | CMOS: 0.000 vs. -0.500 | Helped |
| select k using the boundary predictor instead of k=25 (ablation) | DiffSinger shallow diffusion · PopCS CMOS evaluation | CMOS: 0.000 vs. -0.053 | Helped |
| use 256 channels instead of 128 (ablation) | DiffSinger denoiser · PopCS CMOS evaluation | CMOS: 0.000 vs. -0.071 | Helped |
| use 256 channels instead of 512 (ablation) | DiffSinger denoiser · PopCS CMOS evaluation | CMOS: 0.000 vs. -0.044 | Helped |
| use 20 denoiser layers instead of 10 (ablation) | DiffSinger denoiser · PopCS CMOS evaluation | CMOS: 0.000 vs. -0.293 | Helped |
| use 20 denoiser layers instead of 30 (ablation) | DiffSinger denoiser · PopCS CMOS evaluation | CMOS: 0.000 vs. -0.445 | Helped |
| use shallow diffusion from the auxiliary decoder prediction (ablation) | DiffSpeech · LJSpeech test | MOS: 3.92 ± 0.06 vs. 3.69 ± 0.05 | Helped |
| use shallow diffusion from the auxiliary decoder prediction (ablation) | DiffSpeech · LJSpeech inference | RTF: 29.2% speedup (RTF 0.121 vs. 0.171) | Helped |
| use DiffSinger rather than GAN-Singer | PopCS singing voice synthesis · PopCS test | MOS: 3.85 ± 0.11 vs. 3.74 ± 0.12 | Helped |
| use DiffSinger rather than FFT-Singer | PopCS singing voice synthesis · PopCS test | MOS: 3.85 ± 0.11 vs. 3.67 ± 0.11 | Helped |
Closest papers by meaning
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.90)
- CSSinger: End-to-End Chunkwise Streaming Singing Voice Synthesis System Based on Conditional Variational Autoencoder (AAAI 2025 · similarity 0.88)
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.88)
- StyleSinger: Style Transfer for Out-of-Domain Singing Voice Synthesis (AAAI 2024 · similarity 0.88)
- DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion (AAAI 2024 · similarity 0.88)
- DITTO: Diffusion Inference-Time T-Optimization for Music Generation (ICML 2024 · similarity 0.88)
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.