A technique for dimension reduction of MFCC spectral features for speech recognition

The accuracy of speech recognition systems, to a large extent, depends on the feature sets used for representing the recorded speech data. It has been a continuous process to derive better feature sets for more accurate speech recognition using ASR (Automatic Speech Recognition) systems. Many feature sets and their different combinations have been tried to achieve better accuracy but a feature set providing completely accurate results has not yet been formulated. These large feature sets consume significant amount of memory, together with computing and power requirements and they do not always contribute to improve the recognition rate. The paper investigates the relevance of individual features within the feature sets incorporated in speech recognition systems. The goal is to identify the features that do not contribute significantly in recognition or perhaps causing a fall in the recognition accuracy. The results of the experiments show that about 60% reduction of feature set is feasible with marginal loss of recognition accuracy using our method. The results of the analysis will further be used to formulate better feature sets, smaller than the traditional features with improved accuracy of ASR systems.