@article{1110,
  author = {Hazmoune Samira, Bougamouza Fateh, Mazouzi Smaine, Benmohammed Mohamed},
  title = {Contributions to HMM-based Speech Recognition Systems},
  journal = {International Journal of Computational Linguistics Research},
  year = {2013},
  volume = {4},
  number = {1},
  doi = {},
  url = {http://www.dline.info/jcl/fulltext/v4n1/4.pdf},
  abstract = {In this paper, we propose a new approach based on multiple modeling by Hidden Markov Models (HMM) for isolated word recognition, which aims to maximize word recognition rate by combining several models coming from different start points. Our approach operates on 2 steps; first we create a large set of candidate markovian models for every vocabulary word by changing in their initial models in the Expectation-Maximization (EM) training algorithm, and then we select the best subset of models. The selection of appropriate models to be combined is one of the most difficult but important factors in our approach. For this purpose, we have experimented in our previous work [1] three methods: In the first one, we selected models whose give best individual recognition rates. Secondly, we selected random models. Thirdly, we used a Genetic Algorithm (GA) to select the optimal set of models by maximizing the recognition rate of the group and minimizing the number of selected models. In this paper we propose another new selection method, in which we select models whose maximize the difference between the average of likelihoods of the current class and the average of likelihoods of all others. The performance of the proposed approach will be evaluated by comparing its effectiveness against classical markovian approach.},
}