PURPOSE: Conventional resource-intensive methods for child phonetic development studies are often impractical for sampling and analyzing child vocalizations in sufficient quantity. The purpose of this study was to provide new information on early language development by an automated analysis of child phonetic production using naturalistic recordings. The new approach was evaluated relative to conventional manual transcription methods. Its effectiveness was demonstrated by a case study with 106 children with typical development (TD) ages 8-48 months, 71 children with autism spectrum disorder (ASD) ages 16-48 months, and 49 children with language delay (LD) not related to ASD ages 10-44 months. METHOD: A small digital recorder in the chest pocket of clothing captured full-day natural child vocalizations, which were automatically identified into consonant, vowel, nonspeech, and silence, producing the average count per utterance (ACPU) for consonant and vowel. RESULTS: Clear child utterances were identified with above 72% accuracy. Correlations between machine-estimated and human-transcribed ACPUs were above 0.82. Children with TD produced significantly more consonants and vowels per utterance than did other children. Children with LD produced significantly more consonants but not vowels than did children with ASD. CONCLUSION: The authors provide new information on typical and atypical language development in children with TD, ASD, and LD using an automated computational approach.
PURPOSE: Conventional resource-intensive methods for child phonetic development studies are often impractical for sampling and analyzing child vocalizations in sufficient quantity. The purpose of this study was to provide new information on early language development by an automated analysis of child phonetic production using naturalistic recordings. The new approach was evaluated relative to conventional manual transcription methods. Its effectiveness was demonstrated by a case study with 106 children with typical development (TD) ages 8-48 months, 71 children with autism spectrum disorder (ASD) ages 16-48 months, and 49 children with language delay (LD) not related to ASD ages 10-44 months. METHOD: A small digital recorder in the chest pocket of clothing captured full-day natural child vocalizations, which were automatically identified into consonant, vowel, nonspeech, and silence, producing the average count per utterance (ACPU) for consonant and vowel. RESULTS: Clear child utterances were identified with above 72% accuracy. Correlations between machine-estimated and human-transcribed ACPUs were above 0.82. Children with TD produced significantly more consonants and vowels per utterance than did other children. Children with LD produced significantly more consonants but not vowels than did children with ASD. CONCLUSION: The authors provide new information on typical and atypical language development in children with TD, ASD, and LD using an automated computational approach.
Authors: Mary E Brushe; John Lynch; Sheena Reilly; Edward Melhuish; Murthy N Mittinty; Sally A Brinkman Journal: BMC Pediatr Date: 2021-05-21 Impact factor: 2.125
Authors: Charles R Greenwood; Alana G Schnitz; Dwight Irvin; Shu Fe Tsai; Judith J Carta Journal: Am J Speech Lang Pathol Date: 2018-05-03 Impact factor: 2.408
Authors: Emily C Thompson; Carlos R Benítez-Barrera; Gina P Angley; Tiffany Woynaroski; Anne Marie Tharpe Journal: J Speech Lang Hear Res Date: 2020-01-22 Impact factor: 2.297