Research map / Architectures and efficiency / RNNs and normalization
Non-Autoregressive Neural 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 |
|---|---|---|---|
| Replace the autoregressive Deep Voice 3 acoustic model with non-autoregressive ParaNet | WaveNet-vocoder TTS · English speech test set | MOS: 4.09 ± 0.26 -> 4.01 ± 0.24 | No clear effect |
| Replace Deep Voice 3 with ParaNet | WaveNet-vocoder TTS · 15-sentence test set | inference latency: 1.12 -> 0.024 seconds; 46.7 times speed-up | Helped |
| Use ParaNet rather than Deep Voice 3 | text-to-spectrogram model · 100-sentence test set, no attention mask | attention errors (total): 37 -> 12 | Helped |
| Use ParaNet rather than Deep Voice 3 | text-to-spectrogram model with attention mask · 100-sentence test set | attention errors (total): 8 -> 6 | Helped |
| Add attention masking at synthesis (ablation) | ParaNet · 100-sentence test set | attention errors (total): 12 -> 6 | Helped |
| Add attention masking at synthesis (ablation) | Deep Voice 3 · 100-sentence test set | attention errors (total): 37 -> 8 | Helped |
| Increase decoder depth from 6 to 17 layers (ablation) | ParaNet · 100-sentence test set, no attention mask | attention errors (total): 26 -> 12 | Helped |
| Increase decoder depth from 6 to 17 layers (ablation) | ParaNet · 100-sentence test set, attention mask | attention errors (total): 16 -> 6 | Helped |
| Remove attention distillation (ablation) | ParaNet · TTS synthesis | speech intelligibility | Hurt |
| Remove positional encoding from the attention blocks (ablation) | ParaNet · TTS synthesis | speech intelligibility | Hurt |
| Replace distilled IAF with WaveVAE | ParaNet TTS · English speech test set | MOS: 3.62 -> 3.31 | Hurt |
| Condition WaveGlow on ParaNet-predicted rather than DV3-predicted mel spectrograms | WaveGlow TTS · English speech test set | MOS: 3.92 -> 3.27 | Hurt |
Closest papers by meaning
- ClariNet: Parallel Wave Generation in End-to-End Text-to-Speech (ICLR 2019 · similarity 0.87)
- Bidirectional Variational Inference for Non-Autoregressive Text-to-Speech (ICLR 2021 · similarity 0.87)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.87)
- FastSpeech: Fast, Robust and Controllable Text to Speech (NeurIPS 2019 · similarity 0.86)
- EfficientTTS: An Efficient and High-Quality Text-to-Speech Architecture (ICML 2021 · similarity 0.86)
- Continuous Autoregressive Modeling with Stochastic Monotonic Alignment for Speech Synthesis (ICLR 2025 · 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.