Research map / Architectures and efficiency / Attention and transformers
Efficient Sequence Transduction by Jointly Predicting Tokens and Durations
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 |
|---|---|---|---|
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · Librispeech test-clean | WER (%): 2.14 -> 2.11 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · Librispeech test-clean | time (s): 256 -> 117 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · Librispeech test-other | WER (%): 5.11 -> 5.16 | No clear effect |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · Librispeech test-other | time (s): 244 -> 115 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · Spanish Callhome | WER (%): 19.84 -> 18.73 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · German MLS test | time (s): 558 -> 198 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · MUST-C V2 Test | BLEU (%): 23.21 -> 24.47 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · MUST-C V2 Test | time (s): 218 -> 96 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · SLURP | intent accuracy: 88.53 -> 90.07 | Helped |
| Add joint token-and-duration prediction with durations up to 8 | Conformer-Large transducer · TTS repeated digits | WER%: 64.62 -> 5.78 | Helped |
| Train with sampled conventional transducer loss at probability ω = 0.1 | TDT ASR · Librispeech test-clean, batch=4 | WER (%): RNNT 2.13; TDT 0-2 2.10; TDT 0-4 2.15; TDT 0-6 2.10; TDT 0-8 2.13 | Mixed |
| Increase the maximum supported duration from 2 to 8 (ablation) | TDT · TTS repeated digits | WER%: 12.59 -> 9.35 -> 6.12 -> 5.78 | Helped |
Closest papers by meaning
- MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer (ICLR 2025 · similarity 0.84)
- HAINAN: Fast and Accurate Transducer for Hybrid-Autoregressive ASR (ICLR 2025 · similarity 0.84)
- CLaM-TTS: Improving Neural Codec Language Model for Zero-Shot Text-to-Speech (ICLR 2024 · similarity 0.83)
- Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework (ICLR 2024 · similarity 0.83)
- Rethinking Expressivity and Efficiency in Test-Time Training (NeurIPS 2026 · similarity 0.83)
- Aligner-Encoders: Self-Attention Transformers Can Be Self-Transducers (NeurIPS 2024 · similarity 0.83)
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.