Hidden Markov model length optimization for handwriting recognition systems

This paper investigates the use of three different schemes to optimize the number of states of linear left-to-right hidden Markov models (HMM). In the first method, we describe the fixed length modeling scheme where each character model is assigned the same number of states. The second method considered is the Bakis length modeling where the number of model states is set to a given fraction of the average number of observations of the corresponding character. In the third modeling scheme the number of model states is set to a specified quantile of the corresponding character length histogram. This method is called quantile length modeling. A comparison of different length modeling schemes was carried out with a handwriting recognition system using off-line images of cursively handwritten English words from the IAM database. For the fixed length modeling, a recognition rate of 61% was achieved. Using the Bakis or quantile length modeling the word recognition rates could be improved to over 69%.