Applying prediction techniques to phoneme-based AAC systems

It is well documented that people with severe speech and physical impairments (SSPI) often experience literacy difficulties, which hinder them from effectively using orthographic-based AAC systems for communication. To address this problem, phoneme-based AAC systems have been proposed, which enable users to access a set of spoken phonemes and combine phonemes into speech output. In this paper we investigate how prediction techniques can be applied to improve user performance of such systems. We have developed a phoneme-based prediction system, which supports single phoneme prediction and phoneme-based word prediction using statistical language models generated using a crowdsourced AAC-like corpus. We incorporated our prediction system into a hypothetical 12-key reduced phoneme keyboard. A computational experiment showed that our prediction system led to 56.3% average keystroke savings.