Auditory Based Methods for Robust Speech Feature Extraction

A major limitation for the use of automatic speech recognition in many practical applications is its lacking robustness against changing environmental noise. Automatic speech recognition is based on a sequence of speech feature vectors that contain the information relevant for discriminating between different speech sounds. One possible way to increase the robustness of speech recognition systems is to make the feature vectors less sensitive to the changes in environmental noise, while retaining their good discriminative properties. Humans' exceptional ability to recognize speech in noise has inspired the research on robust feature extraction. Conventional feature extraction methods already incorporate some auditory-based concepts. This paper gives an overview of several alternative feature extraction methods that make use of more detailed knowledge of human speech perception. They have generally shown greater robustness in presence of environmental noise compared to the conventional methods. A common characteristic of the alternative methods is the use of the information about spectral peak positions, which is not sensitive to the changes in environmental noise. This is probably the major reason for their noise robustness, rather than the detailed modeling of the processes in the human auditory system.