An improved SVM using predator prey optimization and Hooke-Jeeves method for speech recognition

For automatic speech recognition, speech signal is represented in terms of feature set. Size of the feature set is an important aspect as it affects recognition accuracy, computational time, memory requirement and complexity of the model. Support vector machine has good application prospects for speech recognition; nevertheless, performance of support vector machine is affected by its parameters. In this research work, a hybrid optimization technique is proposed to improve the learning ability of support vector machine and to select the most appropriate feature set. The hybrid technique integrates predator-prey optimization and Hooke-Jeeves method. To deal with mixed type of decision variables, binary predator-prey optimization technique has also been introduced. During the initial phase, search is performed by predator-prey optimization and to further exploit the search, Hooke-Jeeves method is applied. The proposed technique with support vector machine has been implemented to recognize TI-46 isolated word database in clean as well as noisy conditions and self-recorded Hindi numeral database. The experimental results obtained by proposed technique with support vector machine shows improved recognition rate. Furtheimore ROC curve is analysed to verify sensitivity and specificity of results obtained by proposed technique with support vector machine.