Competitive training: a connectionist approach to the discriminative training of hidden Markov models (speech recognition)

摘要

Presents hidden Markov models (HMMs) within a connectionist framework and shows how error back propagation can be used to discriminatively train HMM parameters. The relationship between this competitive training approach and conventional Baum-Welch re-estimation is explored and experimental results presented for its application in ergodic HMM architectures.