Design of a Novel Hybrid Algorithm for Improved Speech Recognition with Support Vector Machines Classifier

Speaker independent speech recognition system has been a challenging field of research since speech is the most basic and natural means of communication. In this work, a speech recognition system is developed for recognizing isolated words in Malayalam. Here we have used two wavelet based techniques namely Discrete Wavelet Transforms (DWT) and Wavelet Packet Decomposition (WPD) for extracting features from speech. The performance of these methods is tested using Support Vector Machines (SVM) as classifier. A recognition accuracy of 85.4 % is obtained using DWT and SVM combination and 83.2% for WPD and SVM combination. A new feature extraction method is proposed which uses the combined features of both DWT and WPD called Discrete Wavelet Packet Decomposition (DWPD). The feature vectors obtained from this hybrid method is also classified using SVM which produced a better recognition accuracy of 87.8%. Keywords—Discrete Wavelet Transforms, Soft Thresholding, Speech Recognition, Support Vector Machines, Wavelet Packet Decomposition.