The Research Of Speaker Recognition Based On DTED-FTRLS And BP Neural Network

Speaker recognition is a kind of biometrics.It mainly based on the personality features represented in the speaker's voice as feature parameters,so as to utilize various model methods such as dynamic time warping(DTW),vector quantization(VQ),hidden Markov(HMM),Gaussian Mixture Model(GMM)and Artificial Neural Network(ANN).This dissertation is the text-independent speaker recognition system based on improved double-threshold endpoint detection,FTRLS filtering algorithm and BP neural network,the main research is as follows:(1)Improved double-threshold endpoint detection method.This dissertation introduces the algorithm steps of two traditional double-threshold endpoint detection firstly,and verifies that the first endpoint detection scheme has syllable missing detection.The reason may be that muting or pause time is too long,so it is mistakenly considered as the detection end.The second endpoint detection scheme is affected by the burst noise quite seriously.In order to avoid the shortcomings of these two traditional endpoint detection schemes at the same time,this dissertation presents an improved double-threshold endpoint detection method.Compared with the experimental results,we can see that the improved endpoint detection method is effective for speech by each of the syllable sequence detection and removal of the silent part of the syllable interval,while excluding the noise of a certain burst of interference,greatly improving the recognition effect of the speaker system.(2)This dissertation proposes an improved FTRLS filtering algorithm,which is to find out the amount of large error and accumulate the error,and then make the error feedback to make the algorithm more stable.The simulation results show that the improved algorithm can improve the convergence speed and stability of the algorithm,and effectively reduce the convergence of the noise.(3)Using BP Neural Networks for Speaker Recognition.Input layer of BP neural network has 24 or 36 neurons,hidden layer is 25 neurons,the maximum neurons of the output is 10.Besides,the four activation functions,such as Sigmoid,Tanh,ReLu and Leaky ReLu,were analyzed,and then,the improved endpoint Detection and FTRLS algorithms combined with BP neural network are carried on speaker recognition.Experimental results show that the improved algorithm improves the speaker recognition rate by about 5%,reduces the computational complexity and increases the system stability.