Chinese Dialects Identification Using Attention-Based Deep Neural Networks

Abstract

This paper presents a novel Chinese dialects identification system. We use attention-based deep neural networks (AB-DNN) to obtain the Chinese dialects model as back-end. The front-end fuses identity vector (i-vector) with the global prosodic information as input used to describe the dialectal category information accurately. In the task, five kinds of Chinese dialects including Min, Yue, Wu, Jianghuai, Zhongyuan and standard Mandarin are selected as the identification objects. Experimental results show that 21.1% relative equal error rate (EER) reduction is obtained compared with regular deep neural networks (DNN) and further 14.5% reduction when apply global fusion features. The method based on AB-DNN combined with global fusion features observes 29.2% performance improvement compared to traditional DNN with MFCC.