Multilayer perceptrons generate a posteriori probabilities related to emission probabilities of Hidden Markov Models through Bayes rule. This property is used to improve the discrimination of HMM. Moreover, it gives rise to many statistical interpretations which can be cast in neural architectures for nonlinear prediction and triphone probability estimation.
Improved hidden Markov models for speech recognition through neural network learning
Research report RR-94-016
© EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in Research report RR-94-016 and is available at :
PERMALINK : https://www.eurecom.fr/publication/732