Research On Dialect Classification Based On Convolutional Neural Networks

With the advent of the information age,speech recognition has become one of the most important topics in modern computer science research.The development of several new areas,such as medical products,smart homes,and smart vehicles,relies on speech recognition.At present,there are many kinds of related products in speech recognition,such as medical microphones from University of Science and Technology,squeak speakers,Amazon smart speaker echo,and so on.These products are changing life.Speech recognition technology has developed for a long time,but there are still many difficulties,such as: robustness,noise problems,endpoint detection and Chinese recognition.The current speech recognition system now don't support the dialect,so the study of dialect classification and recognition is valuable.China is a multi-ethnic country with great differences in language culture.Therefore,the classification and recognition of Chinese dialects has always been a difficulty in the development of speech recognition.If it is possible to set up database with various local dialects and classify it by using convolutional neural networks,it will be of great significance for the protection of language resources in China,the study of social language life,and the development of Chinese speech recognition.Recently,the National Language Commission has paid more attention to language research projects.Many large companies(such as China Mobile)support the study of Mandarin and dialect recognition programs.In the field of image and speech recognition,the most commonly used is the Convolutional Neural Network(CNN).CNN is a deep neural network with a convolutional structure.Because of its moderate difficulty in modeling and its excellent performance,it has gradually become a widely adopted structure in deep learning.Therefore,this thesis use CNN to classify the dialect dataset.The dataset includes dialects of 70 county-level cities in Jiangsu Province.Dialect dataset is divided into single word,words,speech recordings,and long conversations according to the tracks.Main research is to use the method of convolution neural network to classify the single word and words in the dialect dataset.For the classification of dialect dataset,the thesis used two methods.The first one uses the one-dimensional convolutional neural network to classify the preprocessed audio signals;the second one uses the two-dimensional convolutional neural network and recurrent neural network to classify the dialect spectrogram.