Classification of Zhuang Dialect combined with Bert and SimAM

Abstract

Language recognition refers to the automatic recognition of the language type of the given speech segments. Zhuang is an important language in Southeast Asia, with more than 16 million native speakers. Due to the scarcity of data resources, there is relatively little research about Zhuang in the field of natural language processing, especially in the area of speech-level exploration. Because of the existence of a variety of dialects in Zhuang, this paper is the first to create a speech dataset for the study of the Zhuang language. In the experiment, it was observed that the output of different encoder layers of the BERT has an impact on classification, so this paper innovatively proposes to integrate the SimAM attention mechanism with the BERT model. The model uses a phone recognizer to extract phone sequences from speech segments. Then the Bert model combined with the integrated SimAM attention mechanism performs deep feature representation learning on the semantic information and word order of the sequence. Finally, the classifier realizes the efficient classification of the Zhuang language. Compared with Baseline Systems, the accuracy of the dialect recognition model proposed in this paper is 91.37%, which fully proves the effectiveness of the model.