Feature analysis and compensation for speaker recognition under stressed condition

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
This thesis documents our investigations on feature analysis and design of compensators for speaker recognition under stressed condition. Any condition that causes a speaker to vary his or her speech production from normal to neutral condition is called stressed speech condition. Stressed speech is induced by emotion, high workload, sleep deprivation, frustration and environmental noise. In stressed condition the characteristics of speech signal are different from that of normal or neutral condition. Due to changes in speech signal characteristics the performance of speaker recognition system may degrade under stressed speech conditions. In this work the problem of speaker recognition under stressed condition has been dealt with three broad approaches. First, six speech feature (mel- frequency cepstral coefficients (MFCC) linear prediction (LP) coefficients, linear predictions cepstral (LPCC), reflection coefficients (RC), arc-sin reflection coefficients (ARC) and log area ratio (LAR)) which are widely used for speaker recognition are analysed for evaluation of their characteristics under stressed condition. Statistical technics such as Probability density function (pdf) characteristics, F-ratio test and Kolmogorov-Smirnov (K-S) test are used for evaluation of speaker discrimination capability of the features under stressed condition. Two classifiers, Vector Quantisation (VQ) classifier and Gaussian Mixture Model (GMM) are used to evaluate the speaker recognition results with different speech features.
Supervisor: Samarendra Dandapat