An Ensemble Learning Method for Dialect Classification

Abstract

Dialect Classification Task is the first step of the Multilingual Automatic Speech Recognition System. Because of the difference of accent between dialects in different regions, the problem of Dialect Classification is a very challenging one. Dialect classification is widely used in information processing, military information retrieval and other fields. Therefore, the study of dialect classification is of great significance. This paper proposes an ensemble learning method for dialect classification. Firstly, the low accuracy of dialect data sets is processed and amplified. Then, three models, GRU, CNN and DNN, are used to classify dialects respectively, and the final dialect types are determined by voting. The accuracy of dialect classification by this method is higher than that of the single model with the best performance, the validity of the model is verified.