Research map / Architectures and efficiency / RNNs and normalization
Deep Voice 3: Scaling Text-to-Speech with Convolutional Sequence Learning
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 |
|---|---|---|---|
| Use WaveNet rather than Griffin-Lim waveform synthesis | Deep Voice 3 · Single-speaker English | MOS: $3.62 \pm 0.31$ -> $3.78 \pm 0.30$ | Helped |
| Use WORLD rather than Griffin-Lim waveform synthesis | Deep Voice 3 · VCTK | MOS: $3.01 \pm 0.29$ -> $3.44 \pm 0.32$ | Helped |
| Replace WORLD with Deep Voice 2's WaveNet vocoder and pipeline | VCTK multi-speaker synthesis · VCTK | MOS: $3.44 \pm 0.32$ -> $3.69 \pm 0.23$ | Helped |
| Add inference-time monotonic attention constraint (ablation) | Deep Voice 3 with joint phoneme-and-character input · 100-sentence attention test | Repeat errors: 12 -> 1 | Helped |
| Add inference-time monotonic attention constraint (ablation) | Deep Voice 3 with joint phoneme-and-character input · 100-sentence attention test | Mispronunciation errors: 10 -> 4 | Helped |
| Add inference-time monotonic attention constraint (ablation) | Deep Voice 3 with joint phoneme-and-character input · 100-sentence attention test | Skipped-word errors: 15 -> 3 | Helped |
| Replace dot-product attention with monotonic attention (ablation) | Deep Voice 3 with joint phoneme-and-character input · 100-sentence attention test | Repeat errors: 1 -> 5 | Hurt |
| Replace dot-product attention with monotonic attention (ablation) | Deep Voice 3 with joint phoneme-and-character input · 100-sentence attention test | Skipped-word errors: 3 -> 11 | Hurt |
| Use joint phoneme-and-character input instead of character-only input (ablation) | Deep Voice 3 with inference-time attention constraint · 100-sentence attention test | Mispronunciation errors: 35 -> 4 | Helped |
| Replace Tacotron's recurrent architecture with fully convolutional sequence learning | Single-speaker TTS training · Single-speaker English | Training iteration time: 0.59 seconds -> 0.06 seconds | Helped |
Closest papers by meaning
- Deep Voice: Real-time Neural Text-to-Speech (ICML 2017 · similarity 0.91)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.86)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.85)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.85)
- FocalCodec: Low-Bitrate Speech Coding via Focal Modulation Networks (NeurIPS 2025 · similarity 0.84)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 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.