FPMs have developed a new formalization for the estimation of global a posteriori probabilities from local posteriors estimated by the ANN part of a hybrid system. This formalization is unifying the classical Viterbi trained hybrid system, previous work done on REMAP and a new forward-backward training of hybrid systems.
SU have investigated the Hidden Neural Network architecture, in which a hybrid MLP/HMM is globally optimized using the conditional maximum likelihood criterion. Results on the Phonebook database using this approach are competitive with state-of-the-art results obtained on this database (also in SPRACH). Of particular interest is the result that performing a forward decoding resulted in significantly better results compared with Viterbi decoding.