Speech Intelligibility Prediction Using Binaural Processing for Hearing Loss

As the global issue of hearing loss becomes increasingly severe, developing effective speech intelligibility prediction methods is crucial for improving the performance of hearing aids. However, current methods struggle in noisy environments and overlook individual differences in hearing loss between ears, which impacts prediction accuracy. Therefore, this study proposes a non-intrusive speech intelligibility prediction method that incorporates the binaural processing for hearing loss. The proposed method simulates the multi-stage binaural processing of the outer, middle, and inner ear and integrates binaural cues through an equalization-cancellation model to mitigate masking effects in noisy environments. Key features extracted from speech signals serve as inputs for a hybrid speech intelligibility model combining long short-term memory (LSTM) and light gradient boosting machine (LightGBM) models. The proposed method captures the critical features of speech signals, especially in challenging environments and for different types of hearing loss. Experimental results show that, compared to the baseline system of the second Clarity Prediction Challenge (CPC2) dataset, the proposed method achieves an 8.3% reduction in root mean squared error (RMSE). Notably, the proposed method reduces RMSE by 12.8% when predicting inconsistent hearing loss compared to listeners with consistent hearing levels, confirming the potential of combining hearing loss modeling with binaural processing.