Research map / Probabilistic and scientific modeling / Molecules and proteins
WaveFlow: A Compact Flow-based Model for Raw Audio
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 WaveFlow with h=32 and 128 residual channels instead of Gaussian WaveNet | WaveFlow and Gaussian WaveNet · LJ Speech test | test log-likelihood: 5.059 vs. 5.055 | No clear effect |
| Use a compact WaveFlow (h=16, 64 residual channels) instead of WaveGlow (256 residual channels) | WaveFlow and WaveGlow · LJ Speech speech synthesis | MOS: 4.32 ± 0.08 vs. 4.34 ± 0.11 | No clear effect |
| Increase WaveFlow squeezed height from h=8 to h=64 (ablation) | WaveFlow with 64 residual channels · LJ Speech test | test log-likelihood: 4.935 -> 5.023 | Helped |
| Increase WaveFlow residual channels from 64 to 96 at h=8 (ablation) | WaveFlow · LJ Speech test | test log-likelihood: 4.935 -> 4.977 | Helped |
| Increase WaveFlow flows from 6 to 8 at 96 residual channels and h=8 (ablation) | WaveFlow · LJ Speech test | test log-likelihood: 4.946 -> 4.977 | Helped |
| Use varied height dilations with receptive field 35 instead of all-one dilations with receptive field 17 (ablation) | WaveFlow h=32 · LJ Speech test | test log-likelihood: 4.960 -> 5.055 | Helped |
| Add height reversal permutations after flows (ablation) | WaveFlow h=16, 64 residual channels · LJ Speech test | test log-likelihood: 4.551 -> 4.954 | Helped |
| Use bipartition-and-reverse in addition to height reversals (ablation) | WaveFlow h=16, 64 residual channels · LJ Speech test | test log-likelihood: 4.954 -> 4.971 | Helped |
| Increase WaveFlow residual channels from 64 to 256 at h=16 (ablation) | WaveFlow · LJ Speech speech synthesis | MOS: 4.32 ± 0.08 -> 4.43 ± 0.10 | Helped |
| Increase WaveFlow residual channels from 64 to 256 at h=16 (ablation) | WaveFlow · LJ Speech speech synthesis | synthesis speed over real-time: 42.60× -> 8.42× | Hurt |
| Replace WaveNet with WaveFlow in the Deep Voice 3 vocoder | Deep Voice 3 · Text-to-Speech proprietary dataset | MOS: 4.21 ± 0.08 vs. 4.17 ± 0.09 | No clear effect |
Closest papers by meaning
- DFlow: A Generative Model Combining Denoising AutoEncoder and Normalizing Flow for High Fidelity Waveform Generation (ICML 2024 · similarity 0.91)
- PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation (ICLR 2025 · similarity 0.88)
- RFWave: Multi-band Rectified Flow for Audio Waveform Reconstruction (ICLR 2025 · similarity 0.86)
- UniFlow-Audio: Unified Flow Matching for Audio Generation from Omni-Modalities (NeurIPS 2026 · similarity 0.86)
- BinauralFlow: A Causal and Streamable Approach for High-Quality Binaural Speech Synthesis with Flow Matching Models (ICML 2025 · similarity 0.86)
- Fast Timing-Conditioned Latent Audio Diffusion (ICML 2024 · 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.