Wrapper-based Feature Selection for Classifying Cued Speech Malay Syllables

Abstract

Malay Cued Speech is adapted from English Cued Speech to assist hearing-impaired children in communicating using visual sign language. Malay Cued Speech is created to serve as a supplement of lipreading and allow complete access to spoken language in a purely visual form. Integrating computing technology to Malay Cued Speech offers excellent flexibility of learning and therapy. However, with significant variations of speech signals due to speaker variabilities such as gender, dialects and speaking style could make the task challenging. Without previous understanding on the acoustical properties of the speech, it is difficult to discover the relevant features of the dataset. Besides, irrelevant and redundant features might degrade the classification accuracy due to its large dimension of search space. In this paper, three wrapper-based feature selection using Genetic Algorithm (GA), Particle Swarm Optimization (PSO) and Binary PSO is proposed to discover which features are most useful. The acoustic feature set from 10 native children as configured in Interspeech 2010 Paralinguistic Challenge (IS10) is extracted. Extreme Learning Machine (ELM) is used to classify twenty-two Cued Malay Syllables. The best accuracy is achieved by GA-ELM (72.47%). The optimised features are then fed to a Heterogeneous Ensemble Classifiers (HCE) for further improvement. Radial Basis Function ELM, Polynomial Support Vector Machine and Linear Kernel ELM are constructed for the base classifiers. Multiple combination methods are tested to find diversity among the performances of each base classifiers to attain a significant improvement of the accuracy.