Appliction Of Support Vector Machine In Speaker Identification

Due to its special merits in terms of flexibility, economy, accuracy and safety, speaker recognition technology owns a broad application future in identity verification. At present, the state-of-art speaker identification system performs well under the ideal conditions in experiments. However, the practical results can not meet the requirement of complicated applications in terms of identification accuracy, speed and robustness in noise environment.Support Vector Machine(SVM) is a new classification methodology. It has been proved to be a powerful technique in pattern classification for its good generalization ability. Firstly the principle of SVM is described with details, and then a novel kernel based SVM is proposed and exploited to speaker identification system.The main contributions of the dissertation are as follows:1. Principle component analysis(PCA) approach as feature extraction processing is presented firstly. The kernel technique is combined into PCA to improve the efficiency of the algorithm in unlinear conditions. Kernel principle component analysis(KPCA) is exploited to reduce and denoise the audio feature without losing information..2. A new feature transmittion approach is proposed which works to transmit the audio sequences of variable lengths into feature vector with equal length. On this basis, combining the advantage of Gaussian Mixture Model-Universal Background Model (GMM-UBM) and Support Vector Machine(SVM), a novel kernel function is proposed and used as the kernel function of SVM.3. A binary tree classifier is constructed to implement multi-class identification with two-class classifiers. During the training and testing procedure, the binary tree can remove the repetition of the sample training. So it outperforms one-against-one scheme and one-against-rest scheme, improving speed of training and recognizing.