Biofeedback-Enhanced Treatment for Speech Sound Disorder: Randomized Controlled Trial and Delineation of Sensorimotor Subtypes

点击翻译按钮获取中文摘要

Project Summary/Abstract Children with speech sound disorder show diminished accuracy and intelligibility in spoken communication and may thus be perceived as less capable or intelligent than peers, with negative consequences for both socio- emotional and socioeconomic outcomes [1]–[3]. While most speech errors resolve by the late school-age years, between 2-5% of speakers exhibit residual speech sound disorder (RSSD) that persists through adoles- cence or even adulthood [4], [5], reflecting about 6 million cases in the US. In a series of experimental studies since 2013, our research team has demonstrated that treatment incorporating technologically-enhanced feed- back can improve speech production in individuals with RSSD who have not responded to previous interven- tion [6]–[10]. The primary objective of the parent award (R01 DC017476, “Biofeedback-Enhanced Treatment for Speech Sound Disorder”) is to conduct the first well-powered randomized controlled trial comparing tradi- tional vs biofeedback intervention for the most common type of RSSD, misarticulation of the English /r/ sound. Treatment of RSSD could also be enhanced through the development of tools incorporating artificial intelligence/machine learning (AI/ML). Applications with automated scoring of speech sounds could in principle be used to augment clinician services and achieve higher-intensity practice for faster progress. However, no computerized treatment to date has demonstrated sufficient accuracy for clinical use with children [11]. Existing systems are limited primarily by the fact that publicly available speech corpora have very little representation of either children or individuals with speech impairments. This data scarcity represents a fundamental issue hin- dering advances in clinical applications of automatic speech recognition (ASR) [12]. The proposed sup- plement will address this barrier by modifying and augmenting PERCEPT (Perceptual Error Rating for the Clin- ical Evaluation of Phonetic Targets), an existing corpus of acoustic recordings of child speech assembled through the parent award and the investigators’ previous NIH-funded research since 2013. AI/ML applications of the augmented database are expected to have a twofold scientific impact. First, we anticipate direct benefits for children with RSSD affecting /r/, whose speech samples make up the majority of the current corpus. We will use our database to train a neural network to classify novel child productions containing /r/ as correct or incor- rect. This classifier could then be incorporated into AI tools to increase the efficacy of intervention for children with RSSD. Second, we anticipate that engineers working on the broader problem of ASR for child or clinical speech will be interested in using PERCEPT for model training, especially after the corpus is augmented with more diverse data, as proposed here. We expect to show that acoustic models generated with PERCEPT can improve the performance of currently available open-access speech recognition systems (e.g., Kaldi [13]) in recognition of novel child speech stimuli. The PERCEPT database will be made publicly available through part- nership with PhonBank, an NIH-funded data-sharing platform for speech research (e.g., [14]–[16]).