CNN-based Short-Voice Recognition Technology And Its Application Research

As a biological feature,Phonetics has the characteristics of high activity,high distinction andunrepeatable.Compared with other biometric features(face,fingerprint,finger vein and iris,etc.),speech features has a great advantage in identity authentication.With the rapid development of Internet,network security has become a particularly important link in the present network environment.Major Internet companies have invested a lot of resources for the authentication of net-identity,and achieved certain results,such as the authentication system of Alipay,the voice unlock function of Apple etc.Speech intelligent recognition can be divided into speech recognition and speaker recognition.The former is used to distinguish the content of speech and the latter is used to distinguish the speaker's identity,and it can be applied to the security authentication system.This paper mainly carries out the research of short speech recognition and application based on the convolution neural network.Firstly,based on the problem that the deep neural network training of small data sets is not sufficient,the paper proposes binary data amplification processing.Based on the ASVD data set,this method increases the network recognition rate by 2.1 percentage points,and the convergence process is smoother,ie it is not easy to fall into the local optimum and conducted simulation and performance analysis.Then,based on the ASVD dataset,the dissertation builds a Classified-Convolution Neural Network(C-CNN),and proposes a pattern of continuous segment phrase recognition.Through experiments,the network parameters were optimized.The C-CNN network had a 98.3% recognition rate on the ASVD data set and conducted simulation and performance analysis.Finally,based on the SVD dataset,the paper builds a Aggressive Decision-Convolution Neural Network(AD-CNN).Combined with C-CNN network,this paper proposes a mode of identity authentication based on aggressiveness determination.Through experiments,the network parameters were optimized.The recognition rate of the AD-CNN network on the SVD data set reached 99.4% and conducted simulation and performance analysis.