Research map / Probabilistic and scientific modeling / Molecules and proteins
Blow: a single-scale hyperconditioned flow for non-parallel raw-audio voice conversion
Joan Serrà, Santiago Pascual, Carlos Segura Perales
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 |
|---|---|---|---|
| Use Blow instead of VQ-VAE | Non-parallel voice conversion on VCTK · VCTK test | Spoofing [%]: 65.0 -> 89.3 | Improved |
| Use Blow instead of StarGAN | Non-parallel voice conversion on VCTK · VCTK test | Similarity [%]: 61.8 -> 77.6 | Improved |
| Use Blow instead of StarGAN | Non-parallel voice conversion on VCTK · VCTK test | Naturalness [1--5]: 2.87 -> 2.83 | No clear effect |
| Use Blow instead of Glow-WaveNet | Flow-based voice conversion on VCTK · VCTK test | L [nat/dim]: 4.18 -> 4.45 | Improved |
| Use Blow instead of Glow-WaveNet | Flow-based voice conversion on VCTK · VCTK test | Spoofing [%]: 3.1 -> 89.3 | Improved |
| Replace Blow's 8×12 structure with 3×32 (ablation) | Blow · VCTK, 5 min per speaker, 100 epochs | Spoofing [%]: 66.2 -> 17.2 | Hurt |
| Use 3×32 structure with squeeze of 8 (ablation) | Blow · VCTK, 5 min per speaker, 100 epochs | Spoofing [%]: 66.2 -> 65.7 | No clear effect |
| Use a multi-scale structure (ablation) | Blow · VCTK, 5 min per speaker, 100 epochs | Spoofing [%]: 66.2 -> 3.5 | Hurt |
| Use multi-scale structure with 5×19 flows and squeeze of 4 (ablation) | Blow · VCTK, 5 min per speaker, 100 epochs | Spoofing [%]: 66.2 -> 16.6 | Hurt |
| Replace hyperconditioning with additive conditioning in the coupling network (ablation) | Blow · VCTK, 5 min per speaker, 100 epochs | Spoofing [%]: 66.2 -> 39.5 | Hurt |
| Replace hyperconditioning with additive conditioning before ActNorm (ablation) | Blow · VCTK, 5 min per speaker, 100 epochs | Spoofing [%]: 66.2 -> 22.5 | Hurt |
| Remove data augmentation (ablation) | Blow · VCTK, 5 min per speaker, 100 epochs | Spoofing [%]: 66.2 -> 28.3 | Hurt |
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.
- Glow — Glow: Generative Flow with Invertible 1x1 Convolutions (NeurIPS 2018) · better on 2 of 2 shared results
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
- SelfVC: Voice Conversion With Iterative Refinement using Self Transformations (ICML 2024 · similarity 0.86)
- StableVC: Style Controllable Zero-Shot Voice Conversion with Conditional Flow Matching (AAAI 2025 · similarity 0.86)
- Generative Pre-training for Speech with Flow Matching (ICLR 2024 · similarity 0.86)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.85)
- DFlow: A Generative Model Combining Denoising AutoEncoder and Normalizing Flow for High Fidelity Waveform Generation (ICML 2024 · similarity 0.85)
- DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion (AAAI 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.
