Research map / Image generation / Autoregressive and latent generation
ClariNet: Parallel Wave Generation in End-to-End Text-to-Speech
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 a single Gaussian output distribution instead of a mixture of logistics (ablation) | Autoregressive WaveNet · English speech test audios | Subjective 5-scale MOS: 4.03 ± 0.27 -> 4.40 ± 0.20 | Helped |
| Use a single Gaussian output distribution instead of a mixture of Gaussians (ablation) | Autoregressive WaveNet · English speech test audios | Subjective 5-scale MOS: 4.38 ± 0.22 -> 4.40 ± 0.20 | No clear effect |
| Use a single Gaussian output distribution instead of a mixture of Gaussians (ablation) | Autoregressive WaveNet · English speech test audios | Test CLL: 4.671 -> 4.687 | No clear effect |
| Use forward rather than reverse regularized KL for distillation (ablation) | Student-1 Gaussian IAF · Gaussian WaveNet distillation | Subjective 5-scale MOS: 4.16 ± 0.21 -> 4.12 ± 0.20 | No clear effect |
| Use four flows, distribute WaveNet depth as [10, 10, 10, 30], and reverse sequence direction between flows (ablation) | Student Gaussian IAF · Gaussian WaveNet distillation | Subjective 5-scale MOS: 4.16 ± 0.21 -> 4.22 ± 0.17 | Helped |
| Replace the autoregressive vocoder with a distilled parallel vocoder (ablation) | Text-to-Wave Model · Unseen text | Subjective 5-scale MOS: 4.15 ± 0.25 -> 4.11 ± 0.24 | No clear effect |
| Train the text-to-wave system end-to-end rather than use a separately trained DV3 and WaveNet pipeline with predicted Mel | Text-to-Wave Model · Unseen text | Subjective 5-scale MOS: 3.81 ± 0.26 -> 4.15 ± 0.25 | Helped |
| Train the separately trained pipeline's WaveNet on predicted rather than true Mel (ablation) | DV3 + WaveNet · Unseen text | Subjective 5-scale MOS: 3.73 ± 0.24 -> 3.81 ± 0.26 | Helped |
| Condition WaveNet on hidden representations rather than predicted mel-spectrograms during end-to-end training (ablation) | Text-to-Wave Model · Unseen text | Subjective 5-scale MOS: 3.81 ± 0.26 -> 4.15 ± 0.25 | Helped |
| Train the student IAF with KL divergence alone rather than include the STFT loss (ablation) | Student Gaussian IAF · Speech synthesis | Naturalness | Hurt |
| Use smaller log-scale clipping constant for the Gaussian WaveNet teacher (ablation) | Gaussian autoregressive WaveNet · Hold-out audio | Validation likelihood | Helped |
| Reduce the teacher log-scale clipping constant from -9 to -10 (ablation) | Gaussian autoregressive WaveNet · Hold-out audio | Validation likelihood | No clear effect |
Closest papers by meaning
- Parallel WaveNet: Fast High-Fidelity Speech Synthesis (ICML 2018 · similarity 0.90)
- FloWaveNet : A Generative Flow for Raw Audio (ICML 2019 · similarity 0.88)
- PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation (ICLR 2025 · similarity 0.87)
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.87)
- Non-Autoregressive Neural Text-to-Speech (ICML 2020 · similarity 0.87)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · 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.