Research map / Image generation / Autoregressive and latent generation
PortaSpeech: Portable and High-Quality Generative 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 flow post-net with convolutional post-net (ablation) | PortaSpeech (normal) · LJSpeech | CMOS-Q: -0.441 vs 0.000 | Helped |
| remove mixture alignment (ablation) | PortaSpeech (small) · LJSpeech duration error | word-level duration error (ms): 136.7 -> 96.3 | Helped |
| remove mixture alignment (ablation) | PortaSpeech (small) · LJSpeech duration error | sentence-level duration error (s): 1.84 -> 1.40 | Helped |
| reduce flow-based model from big to small | Flow-based TTS · LJSpeech | MOS-P: 3.71 -> 3.21 | Hurt |
| reduce VAE-based model from big to small | VAE-based TTS · LJSpeech | MOS-P: 3.81 -> 3.72 | No clear effect |
| use VAE instead of flow at big size for audio quality | VAE-based TTS · LJSpeech | MOS-Q: 3.96 -> 3.75 | Hurt |
| shrink PortaSpeech from normal to small | PortaSpeech · LJSpeech | MOS-P: 3.89 -> 3.82 | No clear effect |
| grouped parameter sharing: groups from 1 to 3 (ablation) | PortaSpeech (normal) · LJSpeech | MOS-Q: 3.86 -> 3.91 | Helped |
| grouped parameter sharing: groups from 3 to 12 (ablation) | PortaSpeech (normal) · LJSpeech | MOS-Q: 3.91 -> 3.92 | No clear effect |
| grouped parameter sharing: groups from 1 to 2 (ablation) | PortaSpeech (small) · LJSpeech | MOS-Q: 3.77 -> 3.87 | Helped |
| grouped parameter sharing: groups from 2 to 8 (ablation) | PortaSpeech (small) · LJSpeech | MOS-Q: 3.87 -> 3.89 | No clear effect |
Closest papers by meaning
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.88)
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.87)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.87)
- EmergentTTS-Eval: Evaluating TTS Models on Complex Prosodic, Expressiveness, and Linguistic Challenges Using Model-as-a-Judge (NeurIPS 2025 · similarity 0.87)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.87)
- DiTTo-TTS: Diffusion Transformers for Scalable Text-to-Speech without Domain-Specific Factors (ICLR 2025 · 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.