Automatic Korean dialect identification using wav2vec 2.0 XLS-R*

Previous Korean automatic dialect identification systems have demonstrated limited accuracy, hovering around 60%, and have often relied on traditional machine learning models. A further methodological issue has been the categorization of dialects by administrative districts, a method inconsistent with established linguistic boundaries. This study seeks to overcome these limitations by applying a modern, Transformer-based, end-to-end speech model, wav2vec 2.0 cross-lingual speech representation (XLS-R). The model, pre-trained on approximately 436,000 hours of speech from 128 languages, was fine-tuned for this task. We also re-categorized the dialects into six major linguistic zones: Central, Gyeongnam, Gyeongbuk, Jeonnam, Jeonbuk, and Jeju. Using a dataset of middle-aged and senior male speakers, the XLS-R model was compared against baseline models like Bi-LSTM. The results show a significant performance increase, with the XLS-R model achieving an F1 score of 86.8%—a 14 percentage point improvement over the strongest baseline. While confusion between certain dialects persists, this research validates the effectiveness of applying large-scale, pre-trained models to the nuanced task of dialect identification and underscores the importance of using linguistically-informed categories. Furthermore, the findings contribute to advancing dialect identification technology for applications in speech recognition and forensic science.