Forward search with discontinuous probabilities for online handwriting recognition

Tree structured time-synchronous search methods have proven to be very effective in speech recognition systems, providing a framework in which information about words and language structure can be easily added. It is desirable to apply these same methods to handwriting. However, to apply these search methods it is necessary that every hypothesis can be scored at every point in time, and these scores must accumulate monotonically as the search proceeds. Many of the most efficient character recognizers in online handwriting recognition systems do not provide scores that meet these constraints. If these constraints aren´t met, the search algorithm can no longer prune theories based on their scores lending to an impractical search space size. This paper describes a novel method and protocol that allows discontinuous probability scores of the type produced by many character recognizers to be used with tree structured search methods such as beam and Viterbi search