Research map / Large language models / Language, safety and interpretability
Efficient and Adaptive Simultaneous Speech Translation with Fully Unidirectional Architecture
Biao Fu, Donglei Yu, Minpeng Liao, Chengxi Li, Xinjie Chen, Yidong Chen, Kai Fan, Xiaodong Shi
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 |
|---|---|---|---|
| curate SimulST data with multi-latency semantic chunking instead of using offline ST data | EASiST training data · SimulST data En-De | monotonicity (lower is better): 1.54 -> 1.01 | Improved |
| add policy module and policy loss in Stage III (vs fixed read policy) (ablation) | EASiST · MuST-C En-De tst-COMMON | BLEU vs LAAL | Improved |
| add offline ST loss in Stage III (ablation) | EASiST · MuST-C En-De tst-COMMON | BLEU: 1-2 BLEU at medium and high latency | Improved |
| add Stage II speech-text modality alignment (ablation) | EASiST · MuST-C En-De tst-COMMON | BLEU vs LAAL | Improved |
| add Stage I SimulMT pre-training (ablation) | EASiST · MuST-C En-De tst-COMMON | BLEU vs LAAL | Improved |
| use multi-stage training instead of single-stage (ablation) | EASiST · MuST-C En-De tst-COMMON | BLEU and training hours: 154 -> 33 hours | Improved |
| fine-tune encoder and adapter instead of only encoder (ablation) | EASiST Stage III · MuST-C En-De tst-COMMON | BLEU vs LAAL | Improved |
| fine-tune adapter only instead of encoder+adapter (ablation) | EASiST Stage III · MuST-C En-De tst-COMMON | BLEU vs LAAL | No clear effect |
| also update the LLM during Stage III (ablation) | EASiST Stage III · MuST-C En-De tst-COMMON | BLEU vs LAAL | Hurt |
| raise policy threshold tau from 0.1 to 0.6 (ablation) | EASiST · MuST-C En-De fluency | fluency: 7.51 -> 8.29 | Improved |
| use fully unidirectional architecture with KV cache reuse | EASiST · Inference speed | ms/token: 123.75 (AlignAtt) -> 28.95 | Improved |
Closest papers by meaning
- Divergence-Guided Simultaneous Speech Translation (AAAI 2024 · similarity 0.92)
- SimulMEGA: MoE Routers are Advanced Policy Makers for Simultaneous Speech Translation (NeurIPS 2025 · similarity 0.91)
- High-Fidelity Simultaneous Speech-To-Speech Translation (ICML 2025 · similarity 0.89)
- Large Language Models Are Read/Write Policy-Makers for Simultaneous Generation (AAAI 2025 · similarity 0.87)
- AEQA-NAT : Adaptive End-to-end Quantization Alignment Training Framework for Non-autoregressive Machine Translation (ICML 2025 · similarity 0.86)
- FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing (NeurIPS 2025 · similarity 0.86)
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.
