Studies On The Speech Signals Processing Of The Speaker Recognition System

With the rapid development of the information technology, speech signalprocessing technology has been seeped to each domain in recent years. Speakerrecognition technology is a main part of speech signal processing applications, and it isalso a significant research subject in the field of digital signal processing and speechsignal processing technology at present. Feature extraction is a necessary step of aspeaker recognition system. The selection method and performance optimization offeature extraction is closely related to the feature parameters of speech signal, and themodel parameters of speaker recognition system also have much to do with therecognition performance. Based on the study and research of the forefathers, the mainworks of this thesis according to the above problem are as follows:1. Firstly, this thesis introduces the principle of voice and the speech transmissionmodel. The basic time domain and frequency domain features of speech signal are alsoanalyzed in detail. We use a spectrogram to analysis both them. In addition, thepreprocessing part of speech signal and short-time feature function are also introduced2. Secondly, we introduce three kinds algorithms of speech signal endpointdetection, which are the short-time energy-zcr algorithm, the double thresholdalgorithm and the power spectral entropy algorithm. This paper also shows thesimulation and performance of these three endpoint detection algorithm which arerealized on MATLAB platform, and draws a conclusion: the power spectral entropyalgorithm is the best by contrast. Next then, we makes corresponding improvement atthe shortcomings of the algorithm. And then LPCC and MFCC, which are the twocommon parameters of speech signal feature extraction, are also mentioned in detail.This paper researches about their extract process、parameter characteristic featuresand relating advanced methods such as difference features、 weighing featurecomponents、feature combination.According to the simulation results, this papercompares the MFCC and LPCC feature parameters, draw a important conclusions thatMFCC parameters is better performance than LPCC3. We introduce the Gaussian Mixture Model(GMM) which is a common model training and model match means and the simulation results of the speaker recognitionsystem which of Gauss Mixture Model on the MATLAB platform are shown as well.Moreover, this paper delves into the recognition performance of the GMM models indifferent order、different frame size、and different length of testing speech. Then wetest the recognition rate with the feature parameters after joining difference featureparameters and logarithmic energy. In the end, we try to use a difference featureparameters weighted combination method in view of the feature parameters ofperformance for further improvement and exploration. After that the method is verifiedby experimental results of recognition rate.