Projection On Speech Features Space Improves The Performance Of Speaker Identification

Speaker recognition is task of identifying or verifying who is speaking by analyzing and recognizing speaker-specific information abstracted from speech waves of the speakrr. Gaussion Mixture Model is a popular speaker model for text-independent speaker identification at current. This paper presented some conclusive results for Chinese speaker identification using GMM based on experimental analysis. In feature analysis, the LPCC and MFCC were compared for their performance, the recognition performance was also analyzed with different mixture components and different train speech length and various test speech length and another preprocessing conditionals. As the result, the MFCC was better than LPCC, and a 30 seconds training speech length is enough for model construction, and the error rate become zero when test speech length arrive to 4 seconds for MFCC and 5 seconds for LPCC respectively. Then, this paper presents a projection measure on speech features space and it is applied to speaker identification based on GMM. In the experiment, we used LPCC and MFCC of twelve and fourteen dimensions as speech features, and by analyzing the relative importance of speech features components for speaker identification obtained initial projection modulus. Afterwards, according as the criterion of most low error rate of speaker identification we revised the projection modulus to find optimum projection on speech features space by the algorithm proposed in this paper. The speech features space projected can easily distinguish different sorts and enhance individuality features of speaker in speech features. So, use of the measure improves the recognition performance. In the end of this paper, a novel measure for analyzing subcomponent of feature parameter, named 4S, is presented. This measure is used to analyze and evaluate semantic knowledge and speaker feature information. At last, weanalyzed some property of semantic knowledge and speaker featureinformation, two subcomponents of LPC, MFCC, and LPCC, by DTWexperiment.