Performance analysis of various single channel speech enhancement algorithms for automatic speech recognition

This paper analyzes the performance of various single channel speech enhancement systems when they are applied to automatic speech recognition (ASR) systems as a preprocessor. Until now the researches on speech enhancement algorithms have focused on improving the perceptual quality of speech signal. However, it has not been verified yet whether the improvements of the perceptual quality also increase the speech recognition rate. By investigating several enhancement modules designed for improving the perceptual quality, we analyze the relationship between a speech recognizer and speech enhancement systems. Simulation results show that the decision-directed method and speech absence probability (SAP) estimation proposed for improving perceptual quality influence adverse effects to the speech recognition performance.