Context modeling in a hybrid HMM-neural net speech recognition system

We compare two methods for modeling context in the framework of a hybrid hidden Markov model (HMM)/multilayer perceptron (MLP) speaker-independent continuous speech recognition system. The first method for modeling context is based on the computation of HMM context-dependent observation probabilities using a Bayesian factorization in terms of scaled posterior phone probabilities that are computed with a set of MLPs, one for every relevant context. The second method is based on the use of input features composed of extended multiframe windows of acoustic vectors that include the acoustic information of the current phone as well as various degrees of the acoustic information of the adjacent left and right phones. Experimental results using a hybrid HMM-MLP speaker-independent continuous speech recognition system show that the first approach, based on connectionist context-dependent estimation of observation probabilities, is more efficient in the use of parameters for the same level of recognition performance