Research map / Continual learning / Continual fine-tuning of LLMs
Master-ASR: Achieving Multilingual Scalability and Low-Resource Adaptation in ASR with Modular Learning
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 |
|---|---|---|---|
| replace language-specific or shared weight tuning with modular Specialist Score masks | XLSR-53 multilingual ASR · CommonVoice 51-language multilingual ASR | average CER: 19.48 -> 14.24 | Helped |
| use modular Specialist Score masks instead of a speech adapter | Master-ASR · CommonVoice 51-language multilingual ASR | average CER: 17.35 -> 14.24 | Helped |
| adapt by learning a new Specialist Score combination while preserving pretrained modules | XLSR-53 low-resource ASR · Six low-resource languages, 10-minute training data each | average CER: 27.52 -> 25.26 | Helped |
| further fine-tune the learned mask | Master-ASR low-resource adaptation · Six low-resource languages, 10-minute training data each | average CER: 24.77 -> 24.71 | Helped |
| replace TopK mask generation with the Thres rule (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER: 14.24 -> 18.36 | Hurt |
| replace TopK mask generation with a learned threshold (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER: 14.24 -> 25.21 | Hurt |
| allow backbone weight updates to accommodate learned masks instead of freezing weights (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER: 18.47 -> 14.24 | Helped |
| use iterative rather than random backbone weight updates (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER: 14.41 -> 14.24 | Helped |
| update backbone weights and Specialist Scores together iteratively instead of updating Specialist Scores alone (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER: 27.52 -> 14.24 | Helped |
| increase the mapping-matrix learning-rate multiplier from 0.1 to 10 (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER: 18.39 -> 14.24 | Helped |
| increase the mapping-matrix update interval from beta=1 to beta=5 (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER: 16.21 -> 14.24 | Helped |
| use sparsity ratio t=0.3 rather than other tested ratios (ablation) | Master-ASR · CommonVoice 51-language multilingual ASR | CER | Mixed |
Closest papers by meaning
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.91)
- LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration (AAAI 2025 · similarity 0.91)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (NeurIPS 2024 · similarity 0.89)
- WorldSpeech: A Multilingual Speech Corpus from Around the World (NeurIPS 2026 · similarity 0.89)
- Objective Soups: Multilingual Multi-Task Modeling for Speech Processing (NeurIPS 2025 · similarity 0.89)
- BIGOS V2 Benchmark for Polish ASR: Curated Datasets and Tools for Reproducible Evaluation (NeurIPS 2024 · similarity 0.88)
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.