Out-Of-Vocabulary Rejection For Speech Recognition And Speaker Recognition Based On Supported Vector Machine

In the application of automatic speech recognition, if out-of-vocabulary (OOV) utterances can't be distinguished from in-vocabulary (IV) utterances, incorrect recognition will be concluded, which probably leads to serious results. So it's very important for speech recognition system to reject OOV utterances effectively.An isolated-word speech recognition system based on HMM has been established. Here we mainly studied the techniques of voice activity detector, feature extraction, HMM training, and solved major problems encountered in programming such as data storage, overflow, Multiple observation sequences and so on. Further research will be carried out upon this.We propose a new algorithm for rejection based on post-probability differences and Multi-Layer Perceptron (MLP). First sort the probability output from HMM, calculate the differences between neighbouring ones as the discriminative features into MLP. Then MLP outputs confidence measure. Compared with those existent methods, the increase of computational sum almost can be neglected, and good performance can be achieved.Speaker recognition is a pattern recognition process using personality information in speech signal, and has been widely used in military affairs, judicatory, security, etc.We propose a method for text-independent speaker recognition which combines Support Vector Machines (SVM) with Vector Quantization (VQ). SVM can only classify two classes, how to solve speaker recognition, a problem of classifying multi-class, is the key of our task. After classified by VQ codebooks, the speech signal is put into SVM pairwise classifiers for further processing. Experimental results show that this method not only avoids huge computational burden, but well develops the advantage of SVM.