The earlier identification and treatment of hearing impairment through universal newborn screening, as well as new developments in sensory aid technology such as cochlear implantation, will enable more children with profound hearing impairment (PHI) to produce intelligible speech. Evaluating the effectiveness of these new developments requires accurate, reliable and efficient methods of assessing overall speech intelligibility. The primary goal of this project is to identify a set of parameters that can reliably predict the speech intelligibility of PHI children. This is a necessary first step towards the ultimate goal of developing an automatic means of extracting these measures from a standard speech sample to provide an objective measure for clinical use. This project will take advantage of existing speech samples from a large group of deaf children (N=180), for which some acoustic analyses are nearly complete. These acoustic measures are based on knowledge of segmental and suprasegmental errors made by deaf talkers. Also available, for comparison to any acoustic analysis, are listener intelligibility judgements for these same speech samples. A linear combination of the current acoustic measures provides good, but not clinically acceptable, predictive power. The proposed work will explore two approaches to improving the predictive relation between acoustic measurements and intelligibility. First, the set of acoustic-phonetic measurements will be expanded to include a wider variety of speech characteristics of deaf talkers. Second, from other speech technologies we will adapt measures that reflect gross temporal characteristics (envelope-spectra in the Speech Transmission Index) and spectral characteristics (cepstral coefficients) of the speech signal. Both linear regression and non- linear (e.g., neural net) approaches to combining these measures for maximum prediction will be explored. The groundwork laid by this project will be the basis for an application to develop automatic speech intelligibility assessment software that can be used to determine the efficacy of instructional methods, sensory aids or speech training programs for PHI children.
