Comparison of linear based feature transformations to improve speech recognition performance

In automatic speech recognition system a diagonal GMM based CDHMM modeling is commonly used. So there is a need to use reasonable feature transformation to decorrelate input feature vectors to satisfy diagonal GMM assumption. In this paper, we introduce the utilization of the several supervised linear feature transformation in speech recognition tasks. Specially each of these methods has particular projection properties. We show that the proposed OLPP based feature transformation method with preserving local properties of feature vectors in the projected space has the best performance based on our experiment on Persian speech database FARSDAT. Also we has introduced a novel class labeling method to use the supervised feature transformation. Overall system, compared to the baseline features, achieved an error rate reduction of 22.2% on clean condition.