Phone recognition using hidden Markov models

A phone recognizer based on hidden Markov models is described. For each phone a three-state Markov model is evaluated using the Baum-Welch learning algorithm. Phone models are then connected in a parallel concatenation, representing the phone recognizer. For recognition of a sentence the Viterbi algorithm is used. The recognizer was tested on a database of 50 Slovene sentences spoken by a female speaker.