Reproduce “Two-Stage Adaptation for Non-Normative Speech Recognition: Revisiting Speaker-Independent Initialization for Personalization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shan Jiang, Jiawen Qi, Chuanbing Huo, Yingqiang Gao, Qinyu Chen
A two-stage adaptation strategy, consisting of speaker-independent fine-tuning followed by speaker-specific personalization, enhances Automatic Speech Recognition (ASR) performance for non-normative speech. This approach significantly reduced Word Error Rates (WER) compared to direct speaker-specific adaptation, achieving a 10.2% relative improvement for Whisper-Large-v3 on AphasiaBank and 14.1% on UA-Speech, and even larger gains for Qwen3-ASR.
Key insights
- Population-level adaptation (SI-FT) effectively bridges the distributional gap between generic ASR models and non-normative speech.
- SI-FT provides a superior initialization prior for speaker-specific personalization, leading to consistent performance gains across different speech disorders and ASR backbones.
- Out-of-domain performance on typical speech benchmarks is largely preserved after adaptation to conversational aphasic speech, but shows more pronounced degradation after adaptation to word-level dysarthric speech.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.