Interpreting Accent Classification via Attribution-Based Analysis and Clustering of Time-Frequency Patterns

Recently, many studies have achieved promising results in accent classification using convolutional neural networks (CNNs). However, the interpretability of these models remains a critical challenge, as their decision-making processes are often opaque. To address this, we propose a novel framework that employs attribution methods to analyze the importance distribution of time-frequency representations (TFRs) in relation to the model's output scores. By computing and clustering attribution maps across different accent classes, we systematically identify distinctive time-frequency patterns that characterize each accent. This approach not only enhances the interpretability of CNN-based accent classification but also provides phonetic insights into the discriminative features learned by the model. The results show the structured distribution of the model's attention time-frequency patterns, which reveal the acoustic feature differences in prosody, syllable structure, and phonation habits across accents.