Using articulatory features for speech recognition

Conventional speech recognition draws on the “beads-on-a-string” approach for representing speech; it is modeled as a series of discrete non-overlapping units. Contextual influences cause these models difficulties because the context dependent effects in speech are not explicitly modeled. This paper examines a new type of modeling paradigm that breaks away from traditional segmental-unit modeling and attempts to represent speech as a continuous, smooth articulation process where contextual influences are the result of articulators moving from one configuration to the next. The new modeling paradigm and its implementation are described. Results obtained with the new model show it to perform competitively with conventional models