EEG-based automated evaluation of automotive sound quality using ensemble deep learning

The evaluation of automotive sound quality is of considerable significance for improving driving comfort. However, existing methodologies suffer from notable limitations, including inconsistencies in subjective evaluations and weak correlations between objective metrics and auditory perception. In response to these challenges, an automated evaluation method incorporating electroencephalogram (EEG) signals and ensemble deep learning is proposed herein. Initially, EEG data is acquired from 30 subjects during exposure to 16 automobile sounds with sporty quality. Subsequently, the LSTMS-B model is incorporating Swish activation into LSTM to mitigate gradient vanishing and enhancing Bagging through optimized majority voting, achieving 90.8% accuracy with superior performance over conventional LSTM variants; Furthermore, an innovative ResNet-based regression model is developed to establish the automobile sound-EEG feature mapping, enabling the LSTMS-B model to achieve 89.75% average F1 score in sound quality classification using brain auditory representations while reducing reliance on conventional EEG paradigms. This study develops a novel sound quality evaluation paradigm through deep-ensemble learning integration, where the proposed cross-modal feature mapping method provides a transferable AI framework for interpreting human auditory perception mechanisms.