Research map / Architectures and efficiency / RNNs and normalization
DASpeech: Directed Acyclic Transformer for Fast and High-quality Speech-to-Speech Translation
Qingkai Fang, Yan Zhou, Yang Feng
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 |
|---|---|---|---|
| expect-path training instead of best-path training (lambda=0.5, Lookahead) (ablation) | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 24.45 -> 24.71 | Improved |
| expect-path training instead of best-path training (lambda=0.5, Joint-Viterbi) (ablation) | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 24.84 -> 25.03 | Improved |
| expect-path training instead of best-path training (lambda=1.0, Lookahead) (ablation) | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 24.18 -> 24.41 | Improved |
| expect-path training instead of best-path training (lambda=1.0, Joint-Viterbi) (ablation) | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 24.46 -> 24.80 | Improved |
| end-to-end expect-path finetuning instead of cascading DAT and FastSpeech 2 (lambda=0.5, Lookahead) | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 22.19 -> 24.71 | Improved |
| end-to-end finetuning instead of cascade (lambda=1.0, Joint-Viterbi) | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 23.20 -> 24.80 | Improved |
| smaller graph size lambda=0.5 instead of 1.0 (Joint-Viterbi) (ablation) | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 24.80 -> 25.03 | Improved |
| larger graph size lambda=1.0 instead of 0.5 (Joint-Viterbi) (ablation) | DAT + FastSpeech 2 cascade · CVSS-C Fr-En test | ASR-BLEU: 22.80 -> 23.20 | Improved |
| Joint-Viterbi decoding instead of Lookahead (lambda=0.5) (ablation) | DASpeech · CVSS-C Fr-En test | ASR-BLEU and speedup: 24.71 -> 25.03; 18.53x -> 16.29x | Mixed |
| non-autoregressive DASpeech without knowledge distillation instead of TranSpeech with NPD | DASpeech · CVSS-C Fr-En test | ASR-BLEU: 19.05 -> 25.03 | Improved |
| DASpeech instead of Translatotron 2 in multilingual X-En setting | DASpeech · CVSS-C multilingual X-En test | ASR-BLEU: 8.74 -> 7.43 | Hurt |
| DASpeech instead of S2UT in multilingual X-En setting | DASpeech · CVSS-C multilingual X-En test | ASR-BLEU: 5.15 -> 7.43 | Improved |
Closest papers by meaning
- High-Fidelity Simultaneous Speech-To-Speech Translation (ICML 2025 · similarity 0.87)
- DiffNorm: Self-Supervised Normalization for Non-autoregressive Speech-to-speech Translation (NeurIPS 2024 · similarity 0.87)
- TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling (NeurIPS 2025 · similarity 0.86)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.85)
- DiTAR: Diffusion Transformer Autoregressive Modeling for Speech Generation (ICML 2025 · similarity 0.85)
- FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing (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.
