Research map / Label-efficient and robust learning / Recognition and long-tail
Adversarial Meta Sampling for Multilingual Low-Resource Speech Recognition
Yubei Xiao, Ke Gong, Pan Zhou, Guolin Zheng, Xiaodan Liang, Liang Lin
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 |
|---|---|---|---|
| Add adversarial meta sampling to MML-ASR | MML-ASR · Diversity11, Indo9, and Indo12 target languages | WER (%): 58.29 -> 50.72 (Kyrgyz); 79.66 -> 72.26 (Estonian) | Improved |
| Add adversarial meta sampling to MML-ASR | MML-ASR · IARPA BABEL: Vietnamese, Swahili, and Tamil | CER (%): 45.10 -> 43.35; 36.14 -> 32.19; 50.61 -> 48.56 | Improved |
| Add adversarial meta sampling to MTL-ASR | MTL-ASR · Diversity11, Indo9, and Indo12 | WER (%): 64.90 -> 59.55 (Kyrgyz); 83.70 -> 79.33 (Estonian); 62.55 -> 58.22 (Indo9 Dutch); 84.25 -> 83.90 (Indo9 Kabyle) | Mixed |
| Add attention to the AMS policy network (ablation) | AMS-MAML · Diversity11: Kyrgyz | WER (%): 54.16 -> 50.72 | Improved |
| Use learned adversarial sampling instead of uniform task sampling (ablation) | MML-ASR (MAML) · Diversity11: Kyrgyz and Estonian | WER (%): 58.29 -> 50.72; 79.66 -> 72.26 | Improved |
| Use learned adversarial sampling instead of probability proportional to task quantity (PPQ) (ablation) | MML-ASR (MAML) · Diversity11: Kyrgyz and Estonian | WER (%): 58.95 -> 50.72; 77.26 -> 72.26 | Improved |
| Use learned adversarial sampling instead of query-loss-proportional sampling (PPQL) (ablation) | MML-ASR (MAML) · Diversity11: Kyrgyz and Estonian | WER (%): 54.87 -> 50.72; 74.97 -> 72.26 | Improved |
| Use 80% rather than full target-language training data (ablation) | AMS-MAML · Diversity11: Kyrgyz | WER (%): 50.72 -> 59.02 | Hurt |
| Use 50% rather than full target-language training data (ablation) | AMS-MAML · Diversity11: Kyrgyz | WER (%): 50.72 -> 70.27 | Hurt |
| Use 20% rather than full target-language training data (ablation) | AMS-MAML · Diversity11: Kyrgyz | WER (%): 50.72 -> 87.11 | Hurt |
| Apply AMS to FOMAML | FOMAML · Speech classification, average across D1–D5 | accuracy (%): 37.20 -> 39.63 | Improved |
| Apply AMS to FOMAML | FOMAML · Speech translation: Mongolian, Swedish, and Turkish | BLEU: 0.35 -> 0.36; 0.67 -> 0.70; 1.26 -> 1.45 | Improved |
Closest papers by meaning
- LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration (AAAI 2025 · similarity 0.88)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.88)
- Objective Soups: Multilingual Multi-Task Modeling for Speech Processing (NeurIPS 2025 · similarity 0.87)
- WorldSpeech: A Multilingual Speech Corpus from Around the World (NeurIPS 2026 · similarity 0.86)
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.85)
- OWLS: Scaling Laws for Multilingual Speech Recognition and Translation Models (ICML 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.
