Damped Oscillator Cepstral Coefficients for Robust Speech Recognition.

摘要

This paper presents a new signal-processing technique motivated by the physiology of the human auditory system. In this approach, auditory hair cells are modeled as damped oscillators, which are stimulated by bandlimited speech signals that act as forcing functions. Oscillation synchrony is induced by coupling the forcing functions across the individual bands such that a given oscillator is not only induced by its critical band's forcing function but also by its neighboring functions as well. The damped oscillator model's output is root compressed and cosine transformed to yield a standard cepstral representation. The resulting Synchrony features through Damped Oscillator Cepstral Coefficients (SyDOCC) are used in an Aurora-4 noise- and channel-degraded speech-recognition task, and the results indicate that the proposed feature improved speech-recognition performance in all conditions compared to a baseline using a mel-cepstral feature.