The goal of this research is to understand and improve how individuals utilize continuous word speech recognition (CSR) systems, with a special emphasis on the use of such systems by people with physical disabilities (such as C1-C8 spinal cord injuries, severe repetitive stress injuries, upper extremity amputations, and muscular dystrophy) to enter non-trivial quantities of spontaneous speech. While speech recognition algorithms and accuracy rates continue to improve, identifying and correcting recognition errors remain significant problems plaguing CSR systems. The difficulties are compounded with spontaneous speech input, due to increased recognition errors, or when the users have physical disabilities, due to increased difficulty in identifying and correcting errors. This research addresses these problems by modeling the relationship between user tasks, the difficulties users encounter identifying and correcting errors, feedback accuracy, methodologies used to correct recognition errors once they are identified, and the type and timing of feedback provided by the CSR system. Subsequently, new interaction processes will be defined which increase the effectiveness of CSR systems. The PIs believe increased collaboration between users and CSR systems during the error identification process is critical to increasing the effectiveness of these systems: CSR systems can provide confidence annotations which drive the feedback users receive to help them complete the error identification and error correction processes. This research will result in the development of models that provide unprecedented understanding of the relationship between users and CSR systems, thereby leading to enhanced speech based human machine communication for all users, including those with physical disabilities.
