Perceived conversational competency is an important factor in determining social, academic, and professional success. For persons who require assistive technology to communicate effectively, slow message production can limit advancements in these areas. A number of augmentative techniques designed to accelerate communication rely upon statistical word prediction methods. The proposed study investigates methods of using speech recognition to capture the conversational context provided by the able-bodied conversant to improve word prediction for the augmented conversant. The rapidly changing topic and style of a conversations has a profound impact on the probability of individual words. Techniques such as word recency, word triggers, and discourse domain identification allow statistical language models to exploit shifting context to provide substantially more accurate predictions. In the first stage of the proposed effort, these methods will be applied to transcribed conversations to establish baseline performance measures. To estimate the impact of recognition errors, the studies will be repeated using transcriptions that have been artificially degraded using a carefully tuned speech recognition error simulator. Finally, a commercial speech recognition engine will be integrated into the testing software. The feasibility of this prototype system will be established by applying it to realistic conversations between augmented communicators and able-bodied interlocutors.
The proposed research is targeted at the entire population of augmented communicators, with special emphasis on persons with extremely slow communication rates (less than 5 words per minute) and good language skills. Although the technologies will be developed with an inhouse AAC system, they can be generalized for application to third-party communication devices.
