Estimators Based on Cross-Entropies
Original Publication Date: 1989-Nov-01
Included in the Prior Art Database: 2005-Jan-29
Disclosed is a generalization of the maximum likelihood and related optimization criteria for training and decoding with a speech recognizer. In the maximum likelihood approach to speech recognition (see, e.g.,) ) a maximum probability principle is invoked both for decoding and for training the decoder. During MAP (Maximum A Posteriori) decoding one seeks the a posteriori most probable word sequence. During ML (Maximum Likelihood) training one seeks the parameters of a model which makes the joint likelihood of the training text and the corresponding acoustic information most probable. In the practical implementations of the decoder described in , it was found useful to modify the logarithm of the joint likelihood of words and acoustics by weighting their contributions differently.