The segmentation of the training data has become the most used initialization method for HMM training. The Viterbi reestimation has been widely applied for that purpose. We introduce a new segmentation method, based on the maximization of the Entropic Cohesion Measure (ECM) between segments and observations, which is equivalent to minimize the entropy model. This maximization is carried out by looking for the optimal boundaries between segments of the training utterances. Thus, we obtain an optimal segmentation (in the ECM sense) that achieves similar performance to the Viterbi reestimation with a considerable computational saving.
Cite as: Peinado, A.M., Roman, R., Segura, J.C., Rubio, A.J., Garcia, P., Diaz, J.E. (1991) Entropic training for HMM speech recognition. Proc. 2nd European Conference on Speech Communication and Technology (Eurospeech 1991), 651-654, doi: 10.21437/Eurospeech.1991-160
@inproceedings{peinado91_eurospeech,
author={Antonio M. Peinado and Ramon Roman and Jose C. Segura and Antonio J. Rubio and Pedro Garcia and Jesus E. Diaz},
title={{Entropic training for HMM speech recognition}},
year=1991,
booktitle={Proc. 2nd European Conference on Speech Communication and Technology (Eurospeech 1991)},
pages={651--654},
doi={10.21437/Eurospeech.1991-160}
}