SHAP-Guided explainable feature selection for efficient dialect classification

This study proposes an explainable feature-selection framework, termed SHAP-XFS. It performs dialect detection by integrating Shapley Additive exPlanations (SHAP) with a convolutional neural network (CNN). The objective is to identify the most informative Mel-frequency cepstral coefficients (MFCCs) while improving model efficiency without degrading classification performance. To achieve this, SHAP-XFS : (i) explicitly quantifying the contribution of each MFCC feature using SHAP, (ii) leveraging these attributions for stable feature ranking and pruning, and (iii) systematically evaluating the trade-off between interpretability, accuracy, and computational efficiency. Unlike traditional feature selection methods such as principal component analysis (PCA), least absolute shrinkage and selection operator (LASSO), and Chi-square, provide limited interpretability, the proposed approach enables transparent, reproducible feature selection. Experimental results demonstrate that the SHAP-XFS framework outperforms the baseline on MD3-EN and SLR83 and achieves competitive performance compared to PCA, LASSO, and Chi-square. It significantly reduces feature dimensionality and computational cost, including model parameters, MACs, and memory usage. On the MD3-EN dataset, the proposed framework achieves an accuracy of 82.68% ± 1.78% using only 27 selected MFCC features, improving over the baseline model using 40 MFCC features (77.80% ± 3.34%). On the SLR83 dataset, it achieves 67.09% ± 5.36% using 24 selected MFCC features, outperforming the baseline model using 40 MFCC features (62.08% ± 3.55%). Moreover, the proposed approach reduces model complexity compared to the baseline while maintaining performance comparable to other feature-selection techniques. The selected subsets were highly reproducible, with Jaccard stability scores of 0.9730 for MD3-EN and 0.9698 for SLR83, indicating consistent identification of informative MFCC feature patterns. Statistical tests show that there was no significant difference in performance between SHAP-XFS and PCA-selected features ( $$p > 0.05$$ ). Hence, the proposed SHAP-XFS framework offers a transparent, reliable, and computationally efficient solution for dialect classification.