Feature Selection using Pre-clustering via Affinity Propagation for Speech Classification in Low-resource Languages

Abstract

Speech analysis is an active research field where different feature extraction techniques are studied for solving various issues. Such studies help to improve the time complexity of solutions by understanding necessary clues to select the features. Choosing essential features by removing irrelevant information is a significant step in feature engineering. Perceptual Linear Predictive (PLP) modeling concentrates on understanding the speech signals by focusing on the features perceived at the listener end. They have been used successfully in many speech processing applications. The selection of the order of PLP coefficients for efficient classification of spoken units plays a crucial role in the recognition task. A conventional speech processing system requires a huge training process to develop an Automatic Speech Recognition system. Such systems are efficient for the languages that have enough resources i.e. data. But, low-resource languages especially Asian languages haven't been developed to provide the data sufficient for such tasks. In this context, alternative methods and techniques are encouraged to enhance or optimize the development process with less amount of data. This paper proposes a pre-clustering technique to improve the classification rate with low resources.