Research On Features Of Chinese Dialects Based On Deep Learning

Chinese dialect identification refers to the technology that computers automatically identify the dialect category of a given speech segment.With the development of artificial intelligence,the Chinese dialect identification technology plays an increasingly important role in the multi-language information processing system,especially in China with multiple nations and dialects.A core problem of the Chinese dialect identification is the feature extraction,but the traditional acoustic features are the parameterized representation of speech signal spectral features,often contain speakers,channels,background noises and other redundant information,and increase the back-end modeling and impede the performance improvement of the dialect identification system especially in the short-term speech and highly easily-confused dialect identification tasks.In the face of the aforesaid challenges,it is the current research priority on how to extract more robust features to adapt to the Chinese dialect identification tasks.In recent years,as the deep learning theory has made breakthroughs in the field of machine learning,its non-linear characteristic and ability to extract information deeply and make non-linear modeling have been widely applied.The paper introduces deep learning into feature extraction,tries to inhibit the influences from redundant information in the level of features,strengthen the robustness of the features,and improve the identification performance of the Chinese dialect system.The paper has completed the following tasks:1.It has extracted the deep bottleneck features based on the deep neural network,and applied them in the Chinese dialect identification tasks for the first time.The features complete the non-linear transformation of the multi-frame low-level features training a deep neural network with the bottleneck layer which is related to phonemes,obtains the bottleneck features related to phonemes,effectively inhibit the redundant information unrelated to phonemes,and solves the problem that traditional acoustic features is not robust.2.It has completely constructed the Chinese dialect identification system based on the deep bottleneck features,verified the effectiveness thereof through comparing with the acoustic features,and based on it,discussed about the related parameters of deep bottleneck features and analyzed the reasons for their influences on the system performance.3.Based on the deep bottleneck features of the original single network,it has presented the deep bottleneck features based on hierarchical network.The features firstly purify the acoustic features with a deep automatic encoder,then extract the deep bottleneck features through the second neural network,intensify the abilities of the deep bottleneck features to remove redundant information and improve the robustness of the Chinese dialect feature.