Speech Parameterization for Automatic Speech Recognition in Noisy Conditions

This paper is concerned with increasing the robustness of automatic speech recognition systems (ASR) against additive background noise, by finding speech parameters that are less influenced by changes in acoustic environments than the conventional ones. Inspired by the good robustness of auditory based speech parameterization methods, we compared the processing steps involved with those in the conventional methods. The use of dominant spectral frequencies is believed to be an important reason for the superior robustness of the auditory based methods. A new speech parameterization method is described that is conceptually similar to auditory based methods, while retaining the low computational cost of the conventional methods. It is based on computing the histograms of subband spectrum centroids. Evaluation on an ASR task has shown that the new method outperformed the conventional methods in presence of various background noises.