Research On Speaker Recognition Based On SVM And Deep Learning

With the continuous development of speech recognition technology,speaker recognition technology has received more and more attention as an important method of identity authentication.Traditional speaker recognition technology usually uses MFCC,LPCC,etc.as feature parameters,and the recognition algorithm uses implicit Markov model,vector quantization and Gaussian model,but the speaker recognition technology needs to be further improved in recognition accuracy,identifiable sample size and recognition speed.This thesis mainly studies the following aspects:(1)detailing the model and principle of speaker recognitionThe speech preprocessing stage is studied in detail,and the work of each step in the preprocessing stage is discussed.The specific calculation process of a series of parameters such as MFCC is introduced.Then the mainstream speaker recognition method is studied and four different ways are explored.The speaker recognition model confirms the limitations of mainstream methods.(2)An improved speaker recognition method based on support vector machine and Mel Frequency Cepstrum Coefficient is proposed.In the feature extraction method,the Mel frequency cepstral coefficients are used,and the speech feature parameters are improved.Four improved audio feature parameters are added based on the traditional feature quantities,and then the kernel function types and parameters are analyzed for the SVM model.The experimental results and experimental simulation results show that the improved recognition rate of the speaker recognition system is 21% higher than before.(3)Research on speaker recognition system based on CNN and spectrogramIn this thesis,the speaker's voice information is input as the characteristic parameter,the original information parameter is retained,the speaker's voice signal is processed into a two-dimensional spectrum map,and the format is processed as input,and the spectrum map is processed to obtain different sounds.The pattern is connected to the convolutional neural network to construct a speaker recognition system to test system performance with a recognition rate of 91.2%.