Development of the LEMS speech recognizer: improving performance using feature-sets

This paper discusses the work performed at the Laboratory of Engineering Man-Machine Systems (LEMS) to build a state-of-the-art speech recognizer. In a period of about four years the performance of the speech recognizer has been improved from 83% to 92%, representing a 53% reduction in error rate. These performance gains were obtained by improving the feature-set algorithms and the HMM models. The main change in the feature-set was switching from the once popular LPC-based methods to a novel DFT-based method. The parameterized DFT-based method improved performance and confirmed what has been generally accepted that mel-scale warping is superior for machine speech recognition. Performance gains were also achieved by using the semi-continuous HMM model instead of the fast discrete HMM system. This change appears to offer a fixed 2-3% recognition rate improvement