A improved speech synthesis system utilizing BPSO-based lip feature selection

摘要

To get a higher lipreading recognition result in speech synthesis system driven by visual speech, Binary Particle Swarm Optimization (BPSO) algorithms is used to select the “optimal” lip feature subset. Experiments are carried out based on HMM with 4 states and 16 Gaussian mixture components in a small database for speaker-dependent case. Experiment results show that the integrated discriminate vector after feature selection obtained the information from the geometrical features and the pixel based features. Comparing with feature fusion based on concatenating, the recognition rates with feature selection based on BPSO are improved by as much as 2.42%.