Automating an objective measure of pediatric speech intelligibility

There are many research and clinical settings in which accurate objective measures of speech intelligibility are either necessary or highly desirable. Unfortunately, most objective measures of intelligibility require blinded human listeners to classify speech tokens obtained from the talker whose speech intelligibility is to be measured, a process that can be both costly and time-consuming to carry out. In a research setting, it is often possible to justify the time and cost required, but in a clinical setting this is usually infeasible. The current work describes an effort to develop tools that could be of practical use in a clinical setting. An automatic response scoring system based on 32-Gaussian mixture hidden Markov models was trained on responses of both children with normal hearing and those with hearing loss to the Children's Speech Intelligibility Measure (CSIM) [1]. The system was able to predict a human listener's response out of 12 choices over 60% of the time. Aggregate CSIM scores computed from the ASR system had a high correlation with scores compiled by human listeners (r²=.83). Automatic scoring of utterances from children could allow the CSIM and similar testing procedures to be used more frequently in research and clinical settings.