Research map / Image generation / Autoregressive and latent generation
FPETS: Fully Parallel End-to-End Text-to-Speech System
Dabiao Ma, Zhiba Su, Wenxuan Wang, Yuhao Lu
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 Sin-Cos instead of Gaussian position encoding (ablation) | FPETS alignment · 100 audios | average duration difference: 2.58 -> 0.85 | Improved |
| use trainable instead of fixed position encoding (ablation) | FPETS alignment · 100 audios | average duration difference: 1.96 -> 0.85 | Improved |
| use simple CNN decoder instead of UFANS decoder for alignment learning (ablation) | FPETS alignment · 100 audios | average duration difference: 1.80 -> 0.85 | Improved |
| replace autoregressive decoding with fully parallel generation (vs Tacotron2) (ablation) | FPETS · inference speed | latency (ms): 6157.3 -> 9.9 | Improved |
| fully parallel vs DCTTS (ablation) | FPETS · inference speed | latency (ms): 494.3 -> 9.9 | Improved |
| fully parallel vs Deep Voice 3 (ablation) | FPETS · inference speed | latency (ms): 105.4 -> 9.9 | Improved |
| FPETS vs Tacotron2 quality with Griffin Lim (ablation) | FPETS · Harvard sentences | MOS: 3.51 -> 3.65 | Improved |
| FPETS vs DCTTS quality with Griffin Lim (ablation) | FPETS · Harvard sentences | MOS: 3.55 -> 3.65 | No clear effect |
| FPETS with WaveNet vocoder vs DCTTS (ablation) | FPETS · Harvard sentences | MOS: 3.43 -> 3.27 | Hurt |
| switch vocoder from Griffin Lim to WORLD (ablation) | FPETS · Harvard sentences | MOS: 3.65 -> 3.81 | Improved |
| switch vocoder from Griffin Lim to WaveNet (ablation) | FPETS · Harvard sentences | MOS: 3.65 -> 3.27 | Hurt |
| FPETS vs Tacotron2 mispronunciations (ablation) | FPETS · 100 sentences robustness | mispronunciation count: 5 -> 2 | Improved |
Closest papers by meaning
- EfficientTTS: An Efficient and High-Quality Text-to-Speech Architecture (ICML 2021 · similarity 0.87)
- Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search (NeurIPS 2020 · similarity 0.85)
- A Vector Quantized Approach for Text to Speech Synthesis on Real-World Spontaneous Speech (AAAI 2023 · similarity 0.85)
- TransVIP: Speech to Speech Translation System with Voice and Isochrony Preservation (NeurIPS 2024 · similarity 0.83)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.83)
- Mega-TTS 2: Boosting Prompting Mechanisms for Zero-Shot Speech Synthesis (ICLR 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.
