Research map / Image generation / Autoregressive and latent generation
Toward Complex-Valued Neural Networks for Waveform Generation
Hyung-Seok Oh, Deok-Hyeon Cho, Seung-Bin Kim, Seong-Whan Lee
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 complex-valued networks instead of real-valued (ablation) | MLP GAN · Synthetic complex distribution | JSD (magnitude): 0.018350 -> 0.006548 | Improved |
| use complex-valued networks instead of real-valued (ablation) | MLP GAN · Synthetic complex distribution | JSD (phase): 0.021110 -> 0.003911 | Improved |
| make generator complex-valued (ablation) | ComVo · LibriTTS | UTMOS: 3.6025 -> 3.6452 | Improved |
| make discriminator complex-valued (cMRD) with real generator (ablation) | Vocos-like · LibriTTS | UTMOS: 3.6025 -> 3.5930 | No clear effect |
| make both generator and discriminator complex-valued (ablation) | Vocos-like · LibriTTS | UTMOS: 3.6025 -> 3.6646 | Improved |
| add phase quantization Nq=128 (ablation) | ComVo · LibriTTS | UTMOS: 3.6646 -> 3.6901 | Improved |
| add phase quantization Nq=128 (ablation) | ComVo · LibriTTS | MR-STFT: 0.8435 -> 0.8439 | Hurt |
| increase quantization levels from 128 to 512 (ablation) | ComVo · LibriTTS | UTMOS: 3.6901 -> 3.6412 | Hurt |
| use block-matrix computation instead of native PyTorch complex ops (ablation) | ComVo · LibriTTS | training time (hrs): 183 -> 138 | Improved |
| use block-matrix computation instead of native PyTorch complex ops (ablation) | ComVo · LibriTTS | GPU xRT: 702.26 -> 696.91 | Hurt |
| double real-valued model parameters instead of complex modeling (ablation) | G_R D_R · LibriTTS | UTMOS: 3.6025 -> 3.6164 | No clear effect |
| add phase quantization after first complex conv (ablation) | ComVo · LibriTTS | UTMOS: 3.6646 -> 3.6901 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Vocos — Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis (ICLR 2024) · better on 4 of 4 shared results
- BigVGAN — BigVGAN: A Universal Neural Vocoder with Large-Scale Training (ICLR 2023) · better on 4 of 4 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
- Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis (ICLR 2024 · similarity 0.86)
- Deep Complex Networks (ICLR 2018 · similarity 0.86)
- MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis (NeurIPS 2019 · similarity 0.85)
- Neural Audio Synthesis of Musical Notes with WaveNet Autoencoders (ICML 2017 · similarity 0.85)
- DiffWave: A Versatile Diffusion Model for Audio Synthesis (ICLR 2021 · similarity 0.85)
- DFlow: A Generative Model Combining Denoising AutoEncoder and Normalizing Flow for High Fidelity Waveform Generation (ICML 2024 · similarity 0.83)
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.
