Research map / Architectures and efficiency / Attention and transformers
Speech-T: Transducer for Text to Speech and Beyond
Jiawei Chen, Xu Tan, Yichong Leng, Jin Xu, Guihua Wen, Tao Qin, Tie-Yan Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 Transducer with monotonic alignment instead of attention alignment | TTS · 50 hard sentences | error rate: 32% -> 0% | Improved |
| use Speech-T instead of TransformerTTS for voice quality | TTS · LJSpeech | MOS: 3.77 ± 0.07 -> 3.74 ± 0.07 | No clear effect |
| increase look-ahead k in streaming encoder (ablation) | Speech-T · Streaming TTS | MOS: 3.14 ± 0.19 -> 3.44 ± 0.15 | Improved |
| allow k=2 look-ahead words versus k=0 (ablation) | Speech-T · Streaming TTS | MOS: 3.14 ± 0.19 -> 3.68 ± 0.13 | Improved |
| use lazy forward algorithm instead of GMM-based forward algorithm (ablation) | Speech-T · LJSpeech | CMOS: -1.54 -> 0 | Improved |
| use lazy forward algorithm instead of default Transducer forward algorithm (ablation) | Speech-T · LJSpeech | CMOS: N/A -> 0 | Improved |
| add proposed alignment constraint (ablation) | Speech-T · LJSpeech | CMOS: N/A -> 0 | Improved |
| use proposed alignment constraint instead of handcrafted diagonal constraint (ablation) | Speech-T · LJSpeech | CMOS: -1.33 -> 0 | Improved |
| unify TTS and ASR in one model instead of separate TTS model | Speech-T streaming · LJSpeech | CMOS: -0.09 -> 0 | No clear effect |
| unify TTS and ASR in one model instead of separate ASR model | Speech-T streaming · LJSpeech | PER: 9.7% -> 9.0% | Improved |
Closest papers by meaning
- Efficient Sequence Transduction by Jointly Predicting Tokens and Durations (ICML 2023 · similarity 0.89)
- Aligner-Encoders: Self-Attention Transformers Can Be Self-Transducers (NeurIPS 2024 · similarity 0.87)
- Overcoming Non-monotonicity in Transducer-based Streaming Generation (ICML 2025 · similarity 0.87)
- MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer (ICLR 2025 · similarity 0.86)
- HAINAN: Fast and Accurate Transducer for Hybrid-Autoregressive ASR (ICLR 2025 · similarity 0.86)
- TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling (NeurIPS 2025 · 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.
