SBIR Phase I: Automated Speech Therapy Through Speech Recognition with Error Identification

The broader impact of this Small Business Innovation Research (SBIR) Phase I project is to improve outcomes for individuals with speech sound disorders (SSD) by increasing access to care through technology augmented speech therapy. Approximately 5% of children have a SSD, which places them at risk for reading difficulties and processing challenges into adulthood. Speech therapy services are needed to create changes in speech production skills for improved intelligibility and interaction in functional activities of daily living. However, less than 70% of children receive therapy due to difficulty with accessing care and cost. This project develops artificial intelligence that can recreate the feedback decisions that a speech-language pathologist makes in a live therapy setting. This technology can then be applied to computer- and tablet-delivered speech therapy. Because of the automation, the cost of receiving services can be drastically reduced and children currently without access can get high quality automated therapy at home. This Small Business Innovation Research (SBIR) Phase I project will develop machine learning algorithms capable of speech sound error identification for the purposes of automating speech therapy. Advances in Automatic Speech Recognition technology have led to the exploration of harnessing artificial intelligence as an aspect of technology-assisted endeavors in speech sound error treatment. However, reliability of the existing technology in its application to speech therapy is poor due to inadequate databases of impaired pediatric speech. The objectives of this study are to: 1) Build a database of impaired and accurate productions of the eight English consonants most commonly in error; 2) use this database to develop and train algorithms for identifying specific errors of segmented target phonemes; and 3) complete inter-rater reliability tests of the trained algorithm with decisions from trained speech-language pathologists to ensure adequate agreement of presence of error and error type. By creating a novel training database and algorithm that has the ability to accomplish these types of tasks, artificial intelligence can advance to recreate the error-specific discrimination and feedback decisions that a speech-language pathologist makes in a live therapy setting, serving as a critically needed tool for addressing the barriers to engaging in adequately frequent speech sound production practice sessions. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.