Lip feature selection based on BPSO and SVM

Abstract

In speech synthesis system driven by visual speech, many irrelevant and redundant features will lessen the lipreading recognition result. So it is important to select lip features with stronger discriminate performance. Feature selection algorithm based on binary particle swarm optimization (BPSO) and support vector machines (SVM) is used to select the “optimal” lip feature subset. Feature subset was generated randomly firstly, and then BPSO algorithms searched the feature space guided by the result of SVMs' 10-fold crossover validation. After numbers of iteration, the best fitness feature subset was selected out as the vector of lip feature. Hidden Markov Model (HMM) with 4 states and 16 Gaussian mixture components is designed as a recognizer. Comparing with feature fusion based on concatenating, Experiment results in a small database for speaker-dependent case showed that the recognition rates with feature selection based on BPSO and SVM are improved by as much as 3.89%.