Local Orthogonal Discriminate Bases to Hybrid SVM/Self-adaptive HMM Classifier for Discrete Word Speech Recognition

This research addresses speech recognition with discrete words and speaker independent by using hybrid support vector machine/self-adaptive hidden Markov model (SVM/self-adaptive HMM) classifier. Our proposed method includes two main units: preprocessing unit, and classification unit. In the first unit speech wave is converted into frames and then extracted time-frequency relevant features in the way that maximizing relative entropy of time-frequency energy distribution among frames. In the second unit a one-versus-one self-adaptive HMM calculates words likelihoods to each existed class and then a support vector machine (SVM) takes all class likelihood as an input vector and classifies it into proper class. To validate our proposed method, we test it on our IAUM dataset that contains Persian digits uttered by Persian speakers. The results are compared with similar method using original HMM