Leveraging Multiple Speech Enhancers for Non-Intrusive Intelligibility Prediction for Hearing-Impaired Listeners

Speech intelligibility evaluation for hearing-impaired (HI) listeners is essential for assessing hearing aid performance, but existing listening tests and intrusive metrics such as HASPI require clean reference signals that are often unavailable in real-world conditions. We propose a non-intrusive intelligibility prediction framework that jointly processes noisy and enhanced speech through parallel pathways, enabling robust prediction without reference signals. The model uses cross-attention to explicitly reason over discrepancies between noisy and enhanced representations, serving as a surrogate for intrusive reference information tailored to HI intelligibility prediction. We evaluate three state-of-the-art speech enhancers and show that prediction performance depends strongly on enhancer characteristics, with ensembles of complementary enhancers yielding the best results. To improve cross-dataset generalization, we further introduce a simple 2-clips augmentation strategy that increases listener-specific variability and improves robustness on unseen HI datasets. Experimental results across multiple datasets demonstrate consistent improvements over a strong non-intrusive baseline.