A speech quality metric based on a cochlear model.

Abstract

Noise, distortion, nonlinear signal-processing algorithms, and linear filtering can all affect the sound quality of a hearing aid. The sound quality, in turn, can have a strong impact on the success of the device. The general approach in designing a quality metric is to compare the degraded signal with a clean (unprocessed) reference signal; the comparison generally involves comparing one or more features extracted from the signals. In this presentation, features are extracted from a computationally efficient model of the auditory periphery. The model includes the middle ear, an auditory filter bank, dynamic-range compression, two-tone suppression, and loudness scaling. The first step is a metric for noise, distortion, and nonlinear signal processing. The noise and nonlinear metric focuses on differences in the short-time signal behavior. The second step is a metric that focuses on the changes in the long-term average spectrum caused by linear filtering. The third and final step is to merge the noise and nonlinear metric with the linear filtering metric to produce a composite sound quality metric that can be applied to an arbitrary signal-processing system. The metrics give correlation coefficients better than 0.94 in comparison with quality ratings made by normal-hearing and hearing-impaired listeners.