Automatic speech feature classification for children with cochlear implants

Following pediatric cochlear implantation, it is important to monitor the recipient’s auditory perception skills to (a) determine appropriate therapeutic strategies, and (b) guide adjustment (mapping) of how the CI stimulation is distributed over its channels. One evaluation tool used in clinical and research settings is the OlimSpac [1], which assesses perception at the phonetic feature level by asking children to repeat stimuli that are presented aurally. Errors in the child’s repetition of an item are taken to reflect errors in their perception of the stimulus features. While normally administered by a trained clinician who must score the child’s productions, we present here results from a study involving automatic scoring of productions from 17 pediatric CI patients aged 3;0 to 6;10 years. Taken independently, the automated system was in agreement with human feature classifications over 93% of the time overall (range was 99% for vowel height to 89% for consonant voicing) for stimuli on which human judges were in unanimous agreement. Automatic scoring of utterances from young CI recipients could allow the OlimSpac and similar testing procedures to be used more frequently in a clinical setting, and even administered remotely under appropriate conditions.