Automatic speech recognition using a reduced auditory representation and position-tolerant discrimination

Abstract

The “reduced auditory representation”, or RAR, provides a compact, but comprehensive, description of basilar membrane displacement in a form which is both closely related to overall neural firing statistics and suitable for direct presentation to existing speech recognition algorithms. This paper gives a detailed description of the RAR analysis and defines a preliminary version of a “position-tolerant” distance measure, which could be used with most high resolution representations of speech. The results of recognition experiments are then presented, illustrating the dramatic improvements which can be obtained when the discrimination measure is matched to the data representation.