Research map / Image generation / Autoregressive and latent generation
Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search
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 autoregressive Tacotron 2 with parallel Glow-TTS synthesis | TTS on LJSpeech · LJSpeech test | sampling speed relative to Tacotron 2: 15.7 times faster | Helped |
| Use Glow-TTS instead of Tacotron 2 | Single-speaker TTS · LJSpeech test | 9-scale MOS: 3.88 $\pm$ 0.08 -> 4.01 $\pm$ 0.08 | Helped |
| Use Glow-TTS instead of Tacotron 2 | Multi-speaker TTS · LibriTTS test | 9-scale MOS: 3.35 $\pm$ 0.12 -> 3.45 $\pm$ 0.11 | Mixed |
| Use prior temperature T=0.333 (ablation) | Glow-TTS · LJSpeech test | 9-scale MOS: 3.88 $\pm$ 0.08 -> 4.01 $\pm$ 0.08 | Helped |
| Use prior temperature T=0.500 instead of T=0.333 (ablation) | Glow-TTS · LibriTTS test | 9-scale MOS: 3.20 $\pm$ 0.12 -> 3.31 $\pm$ 0.12 | Helped |
| Use prior temperature T=0.667 instead of T=0.333 (ablation) | Glow-TTS · LibriTTS test | 9-scale MOS: 3.20 $\pm$ 0.12 -> 3.45 $\pm$ 0.11 | Helped |
| Enforce hard monotonic, non-skipping alignments | Glow-TTS · Long utterances from Harry Potter and the Philosopher's Stone | character error rate | Helped |
| Vary the sampled latent noise ε | Glow-TTS · LJSpeech synthesized speech | intonation diversity | Helped |
| Vary temperature T while keeping Gaussian noise fixed | Glow-TTS · LJSpeech synthesized speech | pitch controllability | Helped |
| Scale the predicted durations | Glow-TTS duration predictor · LJSpeech synthesized speech | speaking-rate controllability: 1.25, 1.0, 0.75, and 0.5 | Helped |
| Apply an attention mask (ablation) | DeepVoice 3 · 100 test sentences | total attention errors: 37 -> 8 | Helped |
| Apply an attention mask (ablation) | ParaNet · 100 test sentences | total attention errors: 12 -> 6 | Helped |
Closest papers by meaning
- EfficientTTS: An Efficient and High-Quality Text-to-Speech Architecture (ICML 2021 · similarity 0.90)
- PortaSpeech: Portable and High-Quality Generative Text-to-Speech (NeurIPS 2021 · similarity 0.88)
- Speech Recognition Model Improves Text-to-Speech Synthesis Using Fine-Grained Reward (AAAI 2026 · similarity 0.88)
- Flowtron: an Autoregressive Flow-based Generative Network for Text-to-Speech Synthesis (ICLR 2021 · similarity 0.87)
- Shallow Flow Matching for Coarse-to-Fine Text-to-Speech Synthesis (NeurIPS 2025 · similarity 0.85)
- UniFlow-Audio: Unified Flow Matching for Audio Generation from Omni-Modalities (NeurIPS 2026 · 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.