From lattices of phonemes to sentences: a recurrent neural network approach

The author presents preliminary investigations concerning the use of a sequential neutral network for lexical decoding in continuous speech recognition. They explore the architecture introduced by J.L. Elman (1988) for predicting successive elements of a sequence. This recurrent network admits sequential inputs. This model is based on a multilayer architecture and contains special units, called context units, sensitive to the recent activation history of the network. It is suggested that this model be used for lexical decoding in continuous speech recognition. For that purpose, an extension of Elman´s model is presented in order to treat erroneous sequential inputs and in order to label patterns. It is suggested that the context units be updated considering their previous values and not only the values of the hidden units. Moreover, output units represent words instead of the prediction on the next phoneme. Preliminary experimental results are given