An Evolutionary Programming Based Algorithm for HMM training

In this paper, we propose an evolutionary programming (EP) based algorithm for the training of hidden Markov models (HMMs), which are applied to automatic speech recognition. This algorithm (called the EP algorithm) uses specially designed operators of mutation and selection to find the HMM parameters and the number of states. In order to evaluate the recognition capability of the HMMs trained by the EP algorithm, a computational experiment is carried out using speech samples for 20 words. The results indicate that the proposed EP algorithm is a very promising tool for HMM training the EP-HMM approach achieves an average recognition rate of 97.92% and the training process requires quite a small number of iterations of the algorithm.