State-of-the-art voice recognition technology is based on matching spectral voice patterns (acoustic energy as a function of time and frequency). The signal processing requirements of spectral pattern matching are currently served by special purpose filter banks or digital transforms performed on high performance DSP chips. For several years Votan has been performing an in-depth study of the human auditory system to obtain a better understanding of how signals are processed and speech features extracted by a human being. As part of this research, detailed mathematical models of the physics, chemistry, and neurophysiology of the auditory system have been developed and compared with available experimental data. This research has demonstrated that the signal processing and feature extraction process in a human being are radically different from the spectral pattern approach of current voice recognition systems. The auditory system is extremely sensitive to features not present in the spectral pattern (principally phase and timing features), and conversely is insensitive to features that are prominent in the spectral pattern. These differences are of vital importance for accurate speech recognition. The objective of the proposed Phase I effort is to determine the feasibility of developing a preprocessor for speech recognition which incorporates as accurately as possible a model of the human auditory system. This preprocessor would perform the functions of the outer ear, the middle ear, the inner ear (cochlea), hair cell neural transduction, and neural signal processing in the cochlear nucleus. The output of the preprocessor would be acoustic features suitable for speech recognition systems using either conventional pattern matching techniques or the newer neural net techniques. Anticipated benefits/potential commercial applications - speech recognition is an extremely important area for both commercial and defense applications.
