Training of hidden Markov models for cursive handwritten word recognition

We present a comparison of performances of systems for recognition of handwritten cursive words based on discrete and semi-continuous HMMs. We used lexicon and concatenation of character HMMs to generate word HMM that is matched with input word image. Character models are trained on characters written isolated with simple 16-dimensional low resolution bitmap features. This kind of features enables good visual inspection of the quantization result. Results are given for lexicon of 40 Cyrillic lowercase words. The best recognition rate of 91.5% is achieved with discrete model and PDFs with global distribution parameters. The same system using the 3 best hypotheses gives the recognition rate of 96.7%.