Distortion of Temporal Fine Structure cues in Speech and Analysis of resulting Speech Intelligibility

Abstract

Auditory nerve fiber models provide further insight into the inner workings of the ear and brain. These models have helped us to develop physiologically based speech intelligibility metrics, to assess the difficulty of understanding speech objectively. Several metrics have been developed, but they have been developed using a range of auditory nerve (AN) fiber models. A full comparison of different metrics on even footing should be performed to evaluate the accuracy of their predictions. Speech intelligibility is understood to be dependant on both temporal fine structure and envelope cues in the acoustic speech signal, which are however linked in a way where they are very difficult to split. This makes the evaluation of speech intelligibility metrics tricky, as metrics often aim to analyze mean rate and fine timing information in the auditory nerve representation of the acoustic cues. In this study, a method of phase distortion was developed, with the goal of degrading the fine timing information of a speech signal to the point where only the mean rate representation in the AN is contributing to the speech intelligibility. Also, the neural cross correlation coefficients developed in Heinz & Swaminathan (2009) were adapted from the Zilany & Bruce (2007) auditory nerve model to the Bruce, Erfani & Zilany (2018) AN model.