The Use of Recurrent Neural Networks in Continuous Speech Recognition

This chapter was written in 1994. Further advances have been made such as: context-dependent phone modelling; forward-backward training and adaptation using linear input transformations. This chapter describes a use of recurrent neural networks (i.e., feedback is incorporated in the computation) as an acoustic model for continuous speech recognition. The form of the recurrent neural network is described along with an appropriate parameter estimation procedure. For each frame of acoustic data, the recurrent network generates an estimate of the posterior probability of of the possible phones given the observed acoustic signal. The posteriors are then converted into scaled likelihoods and used as the observation probabilities within a conventional decoding paradigm (e.g., Viterbi decoding). The advantages of using recurrent networks are that they require a small number of parameters and provide a fast decoding capability (relative to conventional, large-vocabulary, HMM systems). wost { if not —ll { —utom—ti™ spee™h re™ognition systems expli™itly or impli™itly ™ompute — @equiv—lentlyD et™FA indi™—ting how well —n input —™ousti™ sign—l m—t™hes — spee™h model of the hypothesised utter—n™eF e fund—ment—l pro˜lem in spee™h re™ognition is how this s™ore m—y ˜e ™omputedD ABSTRACT 1 INTRODUCTION