설명가능한 인공지능을 활용한 대구 부산 방언의 발화 패턴 분석

초록

To enhance the objectivity in analyzing the differences between the Daegu and Busan dialects, this study employed a data-driven approach combined with an interpretation-focused methodology utilizing eXplainable AI. Speech signals were converted into Log Mel-spectrograms and used as input for a Convolutional Neural Network based classification model trained to distinguish between the dialects. Gradient-weighted Class Activation Mapping was then applied to visualize the acoustic features emphasized by the model, thereby improving interpretability regarding dialectal differences. The analysis was conducted under two experimental conditions. In the condition using controlled sentences, clear dialectal differences were observed in specific words. In contrast, with spontaneous speech data, the differences appeared more prominently at the beginning and end of utterances. These results suggest that the well-known dialectal distinctions between Daegu and Busan are not limited to individual lexical items but also manifest in the prosodic features across entire sentences. This demonstrates the effectiveness of interpretable deep learning approaches in the linguistic analysis of regional dialects.

키워드