| dc.description.abstract | Children with speech sound disorders typically improve pronunciation quality by undergoing speech therapy, which must be delivered frequently and with high intensity to be effective. As such, clinic sessions are supplemented with home practice, often under caregiver supervision. However, traditional home practice can grow boring for children due to monotony. Furthermore, practice frequency is limited by caregiver availability, making it difficult for some children to reach therapy dosage. To address these issues, this dissertation presents a novel speech therapy game to increase engagement, and explores automatic pronunciation evaluation techniques to afford children independent practice.
Children with speech sound disorders typically improve pronunciation quality by undergoing speech therapy, which must be delivered frequently and with high intensity to be effective. As such, clinic sessions are supplemented with home practice, often under caregiver supervision. However, traditional home practice can grow boring for children due to monotony. Furthermore, practice frequency is limited by caregiver availability, making it difficult for some children to reach therapy dosage. To address these issues, this dissertation presents a novel speech therapy game to increase engagement, and explores automatic pronunciation evaluation techniques to afford children independent practice.
The therapy game, called Apraxia World, delivers customizable, repetition-based speech therapy while children play through platformer-style levels using typical on-screen tablet controls; children complete in-game speech exercises to collect assets required to progress through the levels. Additionally, Apraxia World provides pronunciation feedback according to an automated pronunciation evaluation system running locally on the tablet. Apraxia World offers two advantages over current commercial and research speech therapy games; first, the game provides extended gameplay to support long therapy treatments; second, it affords some therapy practice independence via automatic pronunciation evaluation, allowing caregivers to lightly supervise instead of directly administer the practice. Pilot testing indicated that children enjoyed the game-based therapy much more than traditional practice and that the exercises did not interfere with gameplay. During a longitudinal study, children made clinically-significant pronunciation improvements while playing Apraxia World at home. Furthermore, children remained engaged in the game-based therapy over the two-month testing period and some even wanted to continue playing post-study.
The second part of the dissertation explores word- and phoneme-level pronunciation verification for child speech therapy applications. Word-level pronunciation verification is accomplished using a child-specific template-matching framework, where an utterance is compared against correctly and incorrectly pronounced examples of the word. This framework identified mispronounced words better than both a standard automated baseline and co-located caregivers. Phoneme-level mispronunciation detection is investigated using a technique from the second-language learning literature: training phoneme-specific classifiers with phonetic posterior features. This method also outperformed the standard baseline, but more significantly, identified mispronunciations better than student clinicians. | en |
