Development and Application of a Transcription Pipeline to Investigate Caregiver–Child Interactions and Clinical Features of Early Childhood Stuttering

Abstract

Stuttering behaviors in early childhood are theorized to be associated with several variables, internal and external to the child. This dissertation investigates many of these variables across three overlapping studies and elaborates on novel methodological approaches for studying young children who stutter and their caregivers, including a digital transcription pipeline that uses automated speech recognition (ASR) technology. Study I revealed that clinical measures of stuttering behavior were largely intercorrelated, but that child-reported communication attitudes were mostly not correlated with stuttering behavior in young children. Study II described the use of ASR transcription systems for communication research, finding improved efficiency with pipelines incorporating ASR transcription of dyadic conversation between children who stutter and primary caregivers compared with human transcription alone, despite substantial word- and speaker-level errors in off-the-shelf ASR systems. Study III found that caregiver question-asking, as well as a composite measure of caregiver demands, significantly increased the momentary likelihood of child stuttering within a 3-second window. Study III also found that caregiver mean response latency and caregiver-child turn ratio were significantly associated with child stuttering at the conversation level. Discussion across studies included alignment of findings with predominant theories—such as the Demands and Capacities Model—and prior evidence, clinical implications, and directions for future translational early stuttering research.