Publication Date
Spring 2025
Degree Type
Thesis
Degree Name
Master of Science (MS)
Department
Communicative Disorders and Sciences
Advisor
Pei-Tzu Tsai; Hahn Koo; Janet Bang
Abstract
Achieving consensus in identifying disfluencies and diagnosing developmental stuttering has been a significant challenge for many clinicians, particularly in young bilingual children, who tend to show increased disfluencies. Artificial intelligence (AI) has the potential to recognize more complex yet systematic patterns that might not be easily observable by even trained clinicians. It also has the potential to bring consistency in the measurement of stuttered speech. This study compares the performance of four AI models, their performance on detecting disfluencies including both stutter-like-disfluencies (SLD) and typical disfluencies (TD) in monolingual and bilingual children's data, then explores whether increasing the quantity of training data would improve model performance in children's data. Results of Study 1 show the one model performed the highest in identifying disfluencies, except for blocks. Results of Study 2 showed that the models had variable performance on identifying TDs and SLDs on bilingual data, whereas it showed consistently better performance on TDs than SLDs on adult data, UCLASS and FluencyBank. Results of Study 3 showed a decrease in performance following the addition of simulated data. Future research is needed to further explore the impact of representative data on model performance and increasing the amount of monolingual and bilingual child data from children who do and do not stutter.
Recommended Citation
Iwakiri, Caryn W. C., "Generalizability of Existing AI Models for Automatic Disfluency Detection in Children's Speech" (2025). Master's Theses. 5651.
DOI: https://doi.org/10.31979/etd.urad-5ej2
https://scholarworks.sjsu.edu/etd_theses/5651
