Research map / Image generation / GANs
HiFi-GAN: Generative Adversarial Networks for Efficient and 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 |
|---|---|---|---|
| Use HiFi-GAN V1 instead of the strongest listed baseline | LJSpeech mel-spectrogram inversion · LJSpeech | MOS: 4.02 ($\pm$0.08) -> 4.36 ($\pm$0.07) | Helped |
| Reduce generator hidden dimension from V1 to V2 | HiFi-GAN · LJSpeech | MOS: 4.36 ($\pm$0.07) -> 4.23 ($\pm$0.07) | Hurt |
| Use the smaller V3 generator configuration rather than V1 | HiFi-GAN · LJSpeech | CPU synthesis speed (kHz): 31.74 -> 296.38 | Helped |
| Use V1 rather than WaveNet (MoL) | HiFi-GAN · VCTK unseen speakers | MOS: 3.52 ($\pm$0.08) -> 3.77 ($\pm$0.07) | Helped |
| Fine-tune using predicted Tacotron2 mel-spectrograms | HiFi-GAN V1 · Tacotron2 end-to-end | MOS: 3.91 ($\pm$0.08) -> 4.18 ($\pm$0.08) | Helped |
| Add MPD (ablation) | HiFi-GAN V3 · LJSpeech ablation | MOS: 2.28 ($\pm$0.09) -> 4.10 ($\pm$0.05) | Helped |
| Add MSD (ablation) | HiFi-GAN V3 · LJSpeech ablation | MOS: 3.74 ($\pm$0.05) -> 4.10 ($\pm$0.05) | Helped |
| Add MRF (ablation) | HiFi-GAN V3 · LJSpeech ablation | MOS: 3.92 ($\pm$0.05) -> 4.10 ($\pm$0.05) | Helped |
| Add mel-spectrogram loss (ablation) | HiFi-GAN V3 · LJSpeech ablation | MOS: 3.25 ($\pm$0.05) -> 4.10 ($\pm$0.05) | Helped |
| Use prime periods rather than powers-of-2 periods (ablation) | HiFi-GAN V3 MPD · LJSpeech ablation | MOS: 3.90 ($\pm$0.05) -> 4.10 ($\pm$0.05) | Helped |
| Add MPD (ablation) | MelGAN · LJSpeech ablation | MOS: 2.88 ($\pm$0.08) -> 3.35 ($\pm$0.07) | Helped |
| Fine-tune using predicted Tacotron2 mel-spectrograms | WaveGlow · Tacotron2 end-to-end | MOS: 3.69 ($\pm$0.08) -> 3.66 ($\pm$0.08) | No clear effect |
Closest papers by meaning
- High Fidelity Speech Synthesis with Adversarial Networks (ICLR 2020 · similarity 0.93)
- GANSynth: Adversarial Neural Audio Synthesis (ICLR 2019 · similarity 0.92)
- MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis (NeurIPS 2019 · similarity 0.90)
- FINALLY: fast and universal speech enhancement with studio-like quality (NeurIPS 2024 · similarity 0.88)
- Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis (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)
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.