Research map / Architectures and efficiency / RNNs and normalization
Deep Voice: Real-time 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 |
|---|---|---|---|
| predict F0 instead of using ground-truth F0 (ablation) | Deep Voice · English speech database | MOS: 3.94±0.26 -> 2.76\pm0.31 | Hurt |
| predict duration and F0 instead of using ground-truth values (ablation) | Deep Voice · English speech database | MOS: 3.94±0.26 -> 2.00\pm0.23 | Hurt |
| use a 20-layer, r=32, s=128 model for 2X real-time inference (ablation) | Deep Voice WaveNet · English speech database | MOS: 3.94±0.26 -> 2.74\pm0.32 | Hurt |
| use a 20-layer, r=64, s=128 model for 1X real-time inference (ablation) | Deep Voice WaveNet · English speech database | MOS: 3.94±0.26 -> 3.35\pm0.31 | Hurt |
| synthesize at 48 kHz instead of 16 kHz (ablation) | Deep Voice WaveNet · English speech database | MOS: 3.94±0.26 -> 3.84\pm0.24 | No clear effect |
| reduce WaveNet depth below 20 layers (ablation) | Deep Voice WaveNet · English speech database | perceptual audio quality | Hurt |
| increase WaveNet depth from 20 to 40 layers (ablation) | Deep Voice WaveNet · English speech database | perceptual audio quality | Helped |
| randomly shift phoneme boundaries by 10-30 milliseconds (ablation) | Deep Voice segmentation annotations · English speech database | audio quality: 10-30 milliseconds | No clear effect |
| use the full original WaveNet feature set instead of the reduced feature set (ablation) | Deep Voice WaveNet · English speech database | perceptual audio quality | No clear effect |
| include extra prosody features such as word and syllable breaks and frame position (ablation) | Deep Voice WaveNet conditioning · English speech database | perceptual audio quality | No clear effect |
| use approximate nonlinearities instead of exact nonlinearities (ablation) | Deep Voice CPU inference · WaveNet inference | performance: roughly 30\% | Helped |
| pin threads to physical cores (ablation) | Deep Voice CPU inference · WaveNet inference | performance: approximately 30\% | Helped |
Closest papers by meaning
- Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning (ICLR 2018 · similarity 0.91)
- Deep Voice 2: Multi-Speaker Neural Text-to-Speech (NeurIPS 2017 · similarity 0.89)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.85)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.85)
- Phoneme-Level Feature Discrepancies: A Key to Detecting Sophisticated Speech Deepfakes (AAAI 2025 · similarity 0.85)
- CLaM-TTS: Improving Neural Codec Language Model for Zero-Shot Text-to-Speech (ICLR 2024 · similarity 0.84)
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.