Research And Implementation Of Speaker Recognition Based On Deep Belief Network

In the new epoch of the Internet and information technology,a great number of voice resources has been expanded to the Internet,so it is definitely useful to process and classify speech.As a way of identity and authentication technology,speaker recognition uses amounts of voice data to distinguish the speaker,and it constitutes one of key techniques in speech signal processing.However,the traditional speaker recognition system still retains insufficient learning,less system complexity,and inefficient big data modeling.When the training data is not sufficient,the complexity of the real model of recognition system is inadequate.Based on the analysis of the advantages and disadvantages of the speaker recognition method,this thesis adopts the deep learning technology and designs a speaker recognition system.The main work of this thesis is as follows:(1)Summarize the features and difficulties of the speaker recognition and feature extraction algorithms,and analyze the advantages and disadvantages of the commonly used speaker recognition models and algorithms.(2)Research on deep learning based speaker recognition framework.The deep learning theory is applied to the traditional speaker recognition system,and the deep belief network is trained by the Restricted Boltzmann Machine and the Back Propagation algorithm,which overcomes the efficiency problem of training directly to the multi-layer network model.(3)In this thesis,we propose an method based on i-vector analysis for speaker recognition in channel conditions.To improve the traditional Gaussian mixture model,a new method is proposed to associate the non-compressed i-vector with deep learning.In the use of uncompressed i-vector form of the speaker recognition method,test the difference of the recognition rate of the traditional method,and the influence of gender on recognition rate.(4)According to the processing flow of speaker recognition,the system structure based on deep learning speaker recognition is constructed,and the core module is designed and simulated.Finally,we test and analyze the performance of various speaker recognition systems.