Research map / Image generation / GANs
Multi-SpectroGAN: High-Diversity and High-Fidelity Spectrogram Generation with Adversarial Style Combination for 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 |
|---|---|---|---|
| Train FastSpeech2 with MSG's conditional adversarial discriminator and feature matching, without reconstruction loss (ablation) | MSG · LJ-speech test | MOS: 3.85 ± 0.03 -> 3.91 ± 0.03 | Helped |
| Remove reconstruction loss from the conditional adversarial MSG (ablation) | MSG · LJ-speech test | MOS: 3.91 ± 0.03 -> 3.14 ± 0.06 | Hurt |
| Add feature matching rather than relying on reconstruction loss (ablation) | MSG · LJ-speech test | MOS: 3.85 ± 0.03 -> 3.91 ± 0.03 | Helped |
| Use frame-level conditional information instead of noise as the discriminator condition (ablation) | MSG · LJ-speech test | MOS | Helped |
| Reduce downsampling size from 3 to 2 (ablation) | MSG · LJ-speech test | CMOS: +0.07 -> 0 | Hurt |
| Increase downsampling size from 3 to 4 (ablation) | MSG · LJ-speech test | convergence: 650k -> 1,000k | Hurt |
| Add ASC using mixup with the same mixing ratio across style and variance (ablation) | MSG · VCTK unseen speakers | MOS: 3.80±0.04 -> 3.82±0.04 | No clear effect |
| Use different mixup ratios for style and each variance (ablation) | MSG+ASC · VCTK seen speakers | MOS: 3.89±0.04 -> 3.90±0.04 | No clear effect |
| Use different mixup ratios for style and each variance (ablation) | MSG+ASC · VCTK unseen speakers | MOS: 3.82±0.04 -> 3.81±0.04 | No clear effect |
| Remove linguistic frame-level condition H_mel from the discriminator (ablation) | MSG · VCTK multi-speaker | MOS | Hurt |
| Remove both speaker style and pitch conditions (ablation) | MSG · VCTK multi-speaker | MOS | Hurt |
| Remove pitch condition p from the discriminator (ablation) | MSG · VCTK multi-speaker | MOS: 3.57± 0.07 -> 3.52± 0.07 | Hurt |
Closest papers by meaning
- MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis (NeurIPS 2019 · similarity 0.88)
- End-to-end Adversarial Text-to-Speech (ICLR 2021 · similarity 0.88)
- HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis (NeurIPS 2020 · similarity 0.87)
- High Fidelity Speech Synthesis with Adversarial Networks (ICLR 2020 · similarity 0.87)
- Metis: A Foundation Speech Generation Model with Masked Generative Pre-training (NeurIPS 2025 · similarity 0.84)
- ELF: Encoding Speaker-Specific Latent Speech Feature for Speech Synthesis (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.