Automatic Screening to Detect ’At Risk’ Child Speech Samples using a Clinical Group Verification framework

Pediatric speech sound disorders (SSD) encompass a wide range of speech production deficits that can interfere with children’s educational growth, social engagement and employment opportunities. Early detection of SSDs can facilitate timely intervention and minimize the potential for life-long adverse effects, but distinguishing between typical and atypical speech production in preschoolers is challenging due to developmental and individual variability in speech acquisition. In this study we apply Gaussian Mixture Models to speech samples from 3- to 6-year-old children, recorded by parents using an iOS app. Speech-language pathologists previously classified the samples as positive (’at risk’ speech, warranting a referral for a speech-language evaluation) or negative (’no risk’ speech). In a series of exploratory analyses, novel distance measures and group scoring techniques are developed which show good subject-level prediction accuracy. Our results provide evidence that it may be feasible to use Speech Processing and Speaker Verification techniques to model and screen speech samples from children for possible speech sound disorders.