Hierarchical partition of the articulatory state space for overlapping-feature based speech recognition

Describes our recent work on improving an overlapping articulatory feature (sub-phonemic) based 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 that is 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.