Third International Conference on Spoken Language Processing (ICSLP 94)
Recently, predictive neural network models (PNNM) have proven successful in various speech recognition tasks. But, they suffer from poor discrimination for acoustically similar speech signals. In this paper, a new discriminative training algorithm based on the minimum-error-rate decision rule is proposed. Experiments on the Korean digits recognition have shown 37.5 % reduction of the number of recognition errors.
Bibliographic reference. Na, KyungMin / Rheem, JaeYeol / Ann, SouGuil (1994): "Minimum-error-rate training of predictive neural network models", In ICSLP-1994, 1479-1482.