Towards a noisy-channel model of dysarthria in speech recognition

Modern automatic speech recognition is ineffective at understanding relatively unintelligible speech caused by neuro-motor disabilities collectively called dysarthria. Since dysarthria is primarily an articulatory phenomenon, we are collecting a database of vocal tract measurements during speech of individuals with cerebral palsy. In this paper, we demonstrate that articulatory knowledge can remove ambiguities in the acoustics of dysarthric speakers by reducing entropy relatively by 18.3%, on average. Furthermore, we demonstrate that dysarthric speech is more precisely portrayed as a noisy-channel distortion of an abstract representation of articulatory goals, rather than as a distortion of non-dysarthric speech. We discuss what implications these results have for our ongoing development of speech systems for dysarthric speakers.