Perceptual and Automatic Evaluations of the Intelligibility of Speech Degraded by Noise Induced Hearing Loss Simulation.

This study aims at comparing perceptual and automatic intelligibility measures on degraded speech. It follows a previous study that designed a novel approach for the automatic prediction of Age-Related Hearing Loss (ARHL) effects on speech intelligibility. In this work, we adapted this approach to a different type of hearing disorder: the Noise Induced Hearing Loss (NIHL), i.e., hearing loss caused by noise exposure at work. Thus, we created a speech corpus made of both isolated words and short sentences pronounced by three speakers (male, female and child) and we simulated different levels of NIHL. A repetition task has been carried out with 60 participants to collect perceptual intelligibility scores. Then, an Automatic Speech Recognition (ASR) system has been designed to predict the perceptual scores of intelligibility. The perceptual evaluation showed similar effects of NIHL simulation on the male, female and child speakers. In addition, the automatic intelligibility measure, based on automatic speech recognition scores, was proven to well predict the effects of the different severity levels of NIHL. Indeed, high correlation coefficients were obtained between the automatic and perceptual intelligibility measures on both speech repetition tasks: 0.94 for isolated words task and 0.97 for sentences task.