Research Of Speaker Recognition Based On GMM And VQ

Speaker Recognition is a biometric authentication technology, which can recognize theidentity of the speaker through analyzing the speaker's voice signal. At present, the speakerrecognition technology is widely used in the field of e-commerce,telephone banking system,access control systems,remote computer logs, etc, which has broad application prospects.Speaker recognition in a quiet environment can get a satisfactory result, however, in a noisyenvironment, the speaker recognition rate will drop dramatically. This involves the speakerrecognition system robustness. Therefore, this article focuses on the speaker recognitionresearch in a noisy environment. The main research work and achievements are as following:(1)The paper analyzes a variety of endpoint detection methods, and simulates the methodsunder different Signal-to-Noise Ratio (SNR), as opposed to the traditional endpoint detectionmethods, in this paper, an improved endpoint detection algorithm based on the differentialenergy three pairsthreshold can get a better detection performance at low SNR.(2)Acoustic feature extraction is the major aspect of the speaker recognition system,speaker recognition in noisy environment is an important topic in the field of speechrecognition. A fusion feature PMFPLP parameter has been proposed in this paper, the featureis used in the robust speaker recognition under low SNR environment. Experiments provedthat the fusion feature not only improve the recognition rate in the noise environment at lowSNR, but simple implementation.(3)Analysis the main technology of the speaker recognition system based on the GMM,study the EM algorithm under the Gaussian mixture model, initialization analysis algorithms,as well as the training and recognition process, also build a speaker recognition system basedon the GMM, and finally simulates and analysis the performance of the difference acousticfeature in the system under different SNR environment.(4) Analysis the main technology of the speaker recognition system based on the VQ,study the LBG algorithm and codebook design principles, also build a speaker recognitionsystem based on the VQ, then use the LBG,GA-K clustering and GA-K-LBG to get the initialcodebook, experimental results show that the GA-K-L algorithm can optimize the syetemperformance. At the same time, integrating feature from acoustic feature MFCC and PLP hasbetter recognition performance, with principal component analysis, the integrated feature canoptimize system performance.