Research map / Image generation / Fast diffusion sampling
Parallel WaveNet: Fast High-Fidelity Speech 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 |
|---|---|---|---|
| Distill WaveNet into a parallel feed-forward model | WaveNet · Professional female speaker, 24kHz, 16-bit PCM | subjective 5-scale MOS: 4.41 ± 0.069 -> 4.41 ± 0.078 | No clear effect |
| Generate samples with the parallel distilled model | WaveNet · NVIDIA P100 GPU, batch size 1 | timesteps/second: 172 timesteps/second -> over 500,000 timesteps/second | Helped |
| Use the 24kHz, 16-bit audio fidelity setup instead of 16kHz, 8-bit audio | WaveNet · Professional female speaker | subjective 5-scale MOS: 4.21 -> 4.41 | Helped |
| Add the perceptual loss to KL and power losses (ablation) | Distilled WaveNet · Preference test versus concatenative baseline | win rate: 60% -> 66% | Helped |
| Add contrastive loss to KL, power, and perceptual losses (ablation) | Distilled WaveNet · Preference test versus concatenative baseline | win rate: 66% -> 65% | No clear effect |
| Add contrastive loss to the distilled model (ablation) | Distilled WaveNet · Speech generation | generated audio noisiness | Helped |
| Use multiple flows instead of a single flow (ablation) | Parallel WaveNet · Professional female speaker | MOS: 4.21 -> 4.41 | Helped |
| Use KL and power losses rather than KL loss alone (ablation) | Distilled WaveNet · Speech generation | naturalness | Helped |
| Train with KL loss alone (ablation) | Distilled WaveNet · Speech generation | naturalness | Hurt |
Closest papers by meaning
- ClariNet: Parallel Wave Generation in End-to-End Text-to-Speech (ICLR 2019 · similarity 0.90)
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.86)
- FocalCodec: Low-Bitrate Speech Coding via Focal Modulation Networks (NeurIPS 2025 · similarity 0.85)
- PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation (ICLR 2025 · similarity 0.85)
- DMOSpeech: Direct Metric Optimization via Distilled Diffusion Model in Zero-Shot Speech Synthesis (ICML 2025 · similarity 0.85)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.85)
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.