Research map / Architectures and efficiency / Attention and transformers
Revisiting End-to-End Speech-to-Text Translation From Scratch
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 |
|---|---|---|---|
| adjust length penalty at inference (ablation) | Baseline Transformer ST · MuST-C En-De tst-COMMON | BLEU: +0.7 | Helped |
| higher dropout (0.2 to 0.4) (ablation) | Baseline Transformer ST · MuST-C En-De tst-COMMON | BLEU: -0.7 | Hurt |
| dropout on raw waveform signals (ablation) | Baseline Transformer ST · MuST-C En-De tst-COMMON | BLEU: -3.5 | Hurt |
| halve training batch size (ablation) | Baseline Transformer ST · MuST-C En-De tst-COMMON | BLEU: -0.5 | Hurt |
| reduce model dimension and attention heads (d_model 512->256, H 8->4) (ablation) | Transformer ST · MuST-C En-De tst-COMMON | BLEU: +0.3 (with fewer params, 20M) | No clear effect |
| enlarge encoder depth to 12 with DS-Init and enlarge FFN to 4096 (ablation) | Transformer ST · MuST-C En-De tst-COMMON | BLEU: +2.1 (19.0 to 21.1) | Helped |
| parameterized distance penalty (PDP) (ablation) | Transformer ST · MuST-C En-De tst-COMMON | BLEU: +1.0 (system 2 to 11) | Helped |
| random initialization of PDP weights instead of ones (ablation) | PDP · MuST-C En-De tst-COMMON | BLEU: -1.2 | Hurt |
| 80-dim filterbank instead of 40-dim (ablation) | Transformer ST with PDP · MuST-C En-De tst-COMMON | BLEU: -0.2 | No clear effect |
| remove delta and delta-delta features (ablation) | Transformer ST with PDP · MuST-C En-De tst-COMMON | BLEU: -0.7 | Hurt |
| small BPE vocabulary (16K to 8K) (ablation) | Transformer ST · MuST-C En-De tst-COMMON | BLEU: 0.0 | No clear effect |
| post-LN changed to pre-LN (ablation) | Deep Transformer ST · MuST-C En-De tst-COMMON | BLEU: -1.2 | Hurt |
Closest papers by meaning
- Regularizing End-to-End Speech Translation with Triangular Decomposition Agreement (AAAI 2022 · similarity 0.91)
- Synchronous Speech Recognition and Speech-to-Text Translation with Interactive Decoding (AAAI 2020 · similarity 0.90)
- High-Fidelity Simultaneous Speech-To-Speech Translation (ICML 2025 · similarity 0.88)
- TransVIP: Speech to Speech Translation System with Voice and Isochrony Preservation (NeurIPS 2024 · similarity 0.86)
- DiffNorm: Self-Supervised Normalization for Non-autoregressive Speech-to-speech Translation (NeurIPS 2024 · similarity 0.86)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · 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.