Use of segmental features in HMM based handwriting recognition

Stochastic pattern recognizers, such as hidden Markov model (HMM) based recognizers, typically use data sample point based features. Examples of point based features for sampled online handwriting may include point position, interpoint vector orientation, and local curvature. In this paper we introduce a method to combine segment based handwriting features with point based features for improved recognition performance. A segment based or segmental feature is a measurement of some characteristic of a contiguous collection of sample points. These segmental feature measurements can be made on any contiguous collection of points, but our approach is to apply segmental features only to certain segments chosen using point based features. Using this method in our system, handwriting recognition error rate is reduced by about 50% over that obtained without segmental features