Perception-based multi-resolution auditory processing of acoustic signals.

摘要

A multi-resolution auditory model is proposed to simulate the spectrotemporal processing of the primary auditory cortex. Inspired by recent physiological findings, the model produces a multi-dimensional representation of cortical activity. Though several nonlinear operations are involved, the inversion of the representation is obtained by applying convex projection technique. A series of psychoacoustical experiments were conducted to estimate the appropriate units for the axes of this auditory model. The "perceptual distance" measure, which was derived from the subjective results, outperforms the independent channel model in threshold prediction tasks. Additionally, a simplified vocal tract model was employed to explore the articulatory equivalence to the cortical axes. This study suggests that both local and global changes in the geometry of the vocal tract result in meaningful changes in the cortical response. The perceptual distance measure, when applied to vowel recognition and timbre quantification, yields better performance than conventional signal processing techniques. Given enough computing power, this perception-based auditory model can be used in many applications like speech recognition, audio coding, and sound identification.