Abstract

We describe our recent work on improving an overlapping articulatory feature (sub-phonemic)b ased speech recognizer with robustness to the requirement of training data. A new decision-tree algorithm is developed and applied to the recognizer design which results in hierarchical partitioning of the articulatory state space. The articulatory states associated with common acoustic correlates, a phenomenon caused by the many-to-one articulation-to-acoustics mapping well known in speech production, are automatically clustered by the decision-tree algorithm. This enables effective prediction of the unseen articulatory states in the training, thereby increasing the recognizer’s robustness. Some preliminary experimental results are provided.