Research On Voiceprint Recognition Based On NCL Incremental Learning

Voiceprint recognition is a hot branch of the biometric identification,and has been widely used in the related fields.In many scenarios,the entire data cannot be gathered at one time,and voice signal may vary with the recording environment and the speaker’s physical and mental state,however,the traditional way of batch type machine learning can hardly handle these scenarios.In this paper we use the incremental learning method to study the voiceprint recognition,which can maintain a high recognition rate in the scenarios above.Incremental learning based on negative correlation learning(NCL)algorithm is used in this paper as the recognition model.Among the existing approaches,the fixed size NCL(FSNCL)forgets learned knowledge easily,growing NCL(GNCL)has low generalization performance,and the selective NCL(SNCL),which based on selective ensemble,is too long.Improvement from two aspects about model training and model selection is mainly made by this paper.Firstly,NCL based on BP network and increased the differences between networks,but it ignores the problems in the training process of network,such as difficult to determine numbers of hidden layer,taking too much training time,and over-fitting.This paper aims at these problems to improve the NCL algorithm by making the network structure adaptive changing,and then combine it with the Bagging algorithm to present a new algorithm.In this paper,the new algorithm is called ANCLBag,it avoids the manual setting the number of hidden nodes,and reduces the number of iterations in the training process,at the same time,increase the differences both in the training set and the network training thus ensures the generalization performance of the ensemble.Next,in the research of incremental learning,this paper draws on the framework of SNCL and replaces NCL with ANCLBag,besides,modifies its model selection method with the method based on clustering and ordering,which takes both the accuracy and the diversity of the model in account.The new algorithm is called as SANCLBag.And it is confirmed in the experiment the generalization performance of the SANCLBag is slightly higher than that of SNCL,however,the time complexity is obviously superior to the latter.Finally,a voiceprint recognition model is built by using the model this paper presents.The model contains the functions of signal preprocessing,feature extraction,incrementaltraining and pattern recognition.The experimental results suggest that the system has high recognition accuracy,and can solve the incremental learning problem effectively.