Author(s)

Xiao Li, Zhe Liu and Zihao Chen

Corresponding Author

Xiao Li

ABSTRACT

This paper presents a dialect species recogn-ition technology based on dilated convolutional neural networks. Dialects as a unique national culture, have rich cultural connotations. If Chinese dialects are to be identified systematically, the dialects must first be classified and summarized to determine the species of the dialects. Three dialects are selected to construct the corpus, and digitize and preprocess the audio data. The features of Mel Frequency Cepstrum Coefficients (MFCC) and FBank are extracted. WaveNet-based convolutional neural network structure is trained to save the best model. The integration of the residual network (ResNet) makes the expressiveness of the network proportional to its depth. Organize classification labels and save the mapping between label classifications and dialects. The experimental results show that the improved accuracy of WaveNet can improve the recognition accuracy to more than 90%, which can be used in dialect accent identification and other fields.

KEYWORDS

Dialect species recognition; waveNet; dilated convolution; feature extraction