Research map / Label-efficient and robust learning / Knowledge distillation
FastSpeech 2: Fast and High-Quality 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 |
|---|---|---|---|
| train on ground-truth mel-spectrograms and add pitch, energy, and more accurate duration inputs | FastSpeech 2 · LJSpeech test | MOS: 3.68 ± 0.09 -> 3.83 ± 0.08 | Helped |
| replace teacher-student training with direct training | FastSpeech 2 · LJSpeech | training time: 53.12 -> 17.02 h | Helped |
| generate waveform directly rather than use a mel-spectrogram vocoder pipeline | FastSpeech 2s · LJSpeech waveform synthesis | inference speed (RTF): 1.95 × 10^{-2} -> 1.80 × 10^{-2} | Helped |
| use MFA duration instead of teacher-model duration | FastSpeech · LJSpeech test | CMOS: 0 -> +0.195 | Helped |
| add energy input (ablation) | FastSpeech 2 · LJSpeech test | CMOS: -0.040 CMOS without energy | Helped |
| add pitch input (ablation) | FastSpeech 2 · LJSpeech test | CMOS: -0.245 CMOS without pitch | Helped |
| add both pitch and energy inputs (ablation) | FastSpeech 2 · LJSpeech test | CMOS: -0.370 CMOS without pitch and energy | Helped |
| add energy input (ablation) | FastSpeech 2s · LJSpeech test | CMOS: -0.160 CMOS without energy | Helped |
| add pitch input (ablation) | FastSpeech 2s · LJSpeech test | CMOS: -1.130 CMOS without pitch | Helped |
| add both pitch and energy inputs (ablation) | FastSpeech 2s · LJSpeech test | CMOS: -1.355 CMOS without pitch and energy | Helped |
| predict pitch in the frequency domain using CWT rather than directly fitting the pitch contour (ablation) | FastSpeech 2 · LJSpeech test | CMOS: 0.185 CMOS drop without CWT | Helped |
| predict pitch in the frequency domain using CWT rather than directly fitting the pitch contour (ablation) | FastSpeech 2s · LJSpeech test | CMOS: 0.201 CMOS drop without CWT | Helped |
Closest papers by meaning
- FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing (NeurIPS 2025 · similarity 0.89)
- 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 0.88)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.88)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.88)
- SyllableLM: Learning Coarse Semantic Units for Speech Language Models (ICLR 2025 · similarity 0.87)
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.