Research On Speaker Recognition Based On Deep Belief Network And Vector Quantization

Speaker recognition is an authentication technique used for identity verification,often referred to as voiceprint recognition.It recognizes the speaker's identity through the personality characteristics of different speaker's voices which is more convenient for user operation.Deep learning has been developed rapidly in recent years.In essence,it is a multilayer nonlinear neural network,which can model complex data relationships.In this paper,deep learning is adopted to optimize the speaker's voice characteristics and improve the system performance in speaker recognition system.The main work is as follows:1.Summarized the advantages and applications of deep learning in speaker recognitionDeep neural network is a complex network model with multiple hidden layers.This paper describes the principles of deep belief network,compares it with other models,further analyzes the advantages of deep neural network in speaker recognition,and summarized the research and application of deep belief networks by researchers at home and abroad.2.Speaker recognition system based on Bottleneck-VQLimited training data leads to inadequate model learning,then affects the recognition performance of the system.The DBN can capture the speaker's personality efficiently in the short speech.This paper chooses traditional MFCC as the input of deep belief networks,then extracts Bottleneck feature,which will be combined with the VQ system,Through experiment analysis,it is proved that the Bottleneck-VQ model can obtain an improvement of 10% over the traditional VQ recognition system when speech duration is limited about 10 s.3.Speaker recognition system based on Auto-Encoder DBN-VQNoisy voice can severely degrade the performance of speaker recognition system.In this paper,DBN is used to construct the deep automatic encoding network,which can effectively filter the noise in speech,and then combined with VQ to construct Auto-Encoder DBN-VQ speaker recognition system.Experimental results demonstrated that Auto-Encoder DBN-VQ based speaker recognition system can obtain an improvement of 15% compared with VQ and GMM systems under noisy conditions.