The Application Of Convolutional Neural Networks In Voiceprint Recognition

As one of the basic biometric technologies, voiceprint recognition is a process for speaker recognition using the features of a voice signal. In comparison with other techniques for identity recognition such as face recognition, fingerprint recognition, palmprint recognition and iris recognition, voiceprint recognition is characteristic of convenience in voice acquisition wherein only a telephone or microphone or other simple equipment is needed, low cost, and high reliability and security. Therefore, it draws much attention of research workers and becomes one of the most active research topics worldwide.Since the acquisition and processing of a voice signal is affected by many factors such as complexity of voice environment, easy imitation, channel changes and so on, there exist many problems to be solved in voiceprint recognition. This thesis focuses on the typical structure of voiceprint recognition, which is based on the structural characteristics of convolution neural network(CNN), and proposes different solutions to the related problems.Firstly, analysis of the advantages and disadvantages of the main algorithms for solving the related problems existed in different parts of a voiceprint recognition system is presented, followed by proposition of an approach for parameter estimation in classical voiceprint recognition using Mel frequency cepstral coefficients as features and Gaussian mixture model for classification, wherein experimental results are presented to verify the effect of parameter values on the system performance.Secondly, deep analysis is given to the structure of a CNN and its two typical operations for convolution and down sampling, and based on this, CNN is proposed to be used for extraction of voice features which are invariant to shift, rotation and scaling operations, follow by proposition of two operations in preprocessing of a voice signal, 1-D and 2-D convolution processing, and presentation of experimental verification of the performance in comparison with the classical system.Finally, feature extraction and classification is combined, using CNN, for voiceprint recognition, in which the spectrograms of a voice signal are calculated and input to a CNN follow by output of the category indexes. The combination makes full use of the advantage of a CNN in solving the difficulty problems of parameter selection and model training in traditional approaches. Experimental results show the feasibility of the proposed method for voiceprint recognition.