KOREAN DIALECT IDENTIFICATION BASED ON INTONATION MODELING

Abstract

Korean dialect identification (K-DID) is a challenging task due to its relatively unexplored field of study, mutual comprehensibility between the dialects, and lack of sufficient Korean dialect datasets available in the past. With large-scaled dialect datasets now available, this paper proposes intonational modeling of the Korean dialects by feeding frame-wise acoustic features on sequential modeling of a neural network. Compared to previous prosodic labeling with syllable-based pitch marking, our approach of intonation modeling is realized with the combination of a set of spectral features, including fundamental frequency, trained on a bidirectional LSTM network with attention mechanism. We believe the attention mechanism enables the detection of dialect-rich segments hidden among the dominant non-dialect segments within the same utterance. We test the networks on different combinations of speaker ages and speech styles. The best performance of the K-DID is achieved with 68.51% in utterance-level accuracy, which surpasses our previous work.