Learning mixed acoustic/articulatory models for disabled speech

This paper argues that automatic speech recognition (ASR) should accommodate dysarthric speech by incorporating knowledge of the production characteristics of these speakers. We describe the acquisition of a new database of dysarthric speech that includes aligned acoustics and articulatory data obtained by electromagnetic articulography. This database is used to train theoretical and empirical models of the vocal tract within ASR which are compared against discriminative models such as neural networks, support vector machines, and conditional random fields. Results show significant improvements in accuracy over the baseline through the use of production knowledge.