Recurrent HMMs and Cursive Handwriting Recognition Graphs

Standard cursive handwriting recognition is based on a language model, mostly a lexicon of possible word hypotheses or character n-grams. The result is a list of word alternatives ranked by confidence. Present-day applications use very large language models, leading to high computational costs and reduced accuracy. For a standard HMM-based word recognition system, a new recurrent HMM approach for very fast lexicon-free recognition will be presented. The evaluation of this model creates a "recognition graph", a compact representation of result alternatives of lexicon-free recognition. This structure is formally identical to results of single character segmentation and recognition. Thus it can be directly evaluated by interpretation algorithms following this process, and can even be merged with these results. In addition, the recognition graph is a basis for further evaluation in terms of word recognition. It allows fast evaluation of word hypotheses, easy integration of various language models like n-grams, and the efficient extraction of lexicon-free n-best result alternatives.