Dialect Species Recognition Based On Wavenet

Under the environment in which the country is fully promoting Mandarin,the use of dialects has gradually decreased.As a unique language form,dialects have rich cultural heritage.Based on the cultural forms of different regions,China has formed local dialects with different forms.How to use deep learning to identify species in various dialects has become the focus of current research.However,most of the existing technologies are aimed at language recognition,and there are few studies on the recognition of Chinese dialect species and the recognition rate needs to be improved.In addition,the boundary between dialects and dialects is relatively vague,and the extraction of dialect features also encountered bottlenecks.In response to the above problems,this article carried out the following work:(1)Collected six types of dialect audio from different channels.Based on the six kinds of experimental data in the data set,the differences between the audios of dialects are analyzed.It is proposed to combine the pitch features in Chinese dialects with artificial acoustic features as the object of feature extraction.(2)Establish a dialect species recognition model based on Wavenet.The dilated causal convolution method is used for convolution operation,and the gating activation mechanism and residual network are introduced to establish a recognition model.Compared with other classification algorithms,the accuracy rate is 1.57% higher than that of Google NET.(3)To solve the problem of model scale,use knowledge distillation technology to compress the model.Transfer the knowledge acquired by the teacher's network to a streamlined model to guide it in network training.Thus,a smaller model is obtained,and the loss of recognition accuracy does not exceed 2%.This method can accelerate model training while ensuring performance.(4)Analysis and design of dialect species identification system.Apply the proposed classification algorithm to the system.According to the detailed program design,the function of the entire system is realized.The test results show that the response time of the system to recognition is stable within 3s,and the recognition rate reaches 98.33%.The dialect species recognition system implemented in this paper can classify the dialect audio in detail,providing strong technical support for the continuation of dialect culture.As a front-end processing technology for dialect recognition,it will pave the way for future text recognition.The model compression technology used at the same time provides convenience for application deployment to mobile devices.