A novel algorithm for HMM word spotting performance evaluation and error analysis
J.N. Marcus · 1992
A hidden Markov model (HMM) wordspotter is described. The emphasis is on the algorithms for HMM scoring and performance evaluation, which offer several advantages over those currently used. These advantages include the ability to: determine both the beginning and ending points of a spotted word, generate a smooth receiver operating characteristic (ROC) in a computationally efficient manner, and compare word spotters on the same task using a nonparametric significance test.>