Implementation of an autoassociative recurrent neural network for speech recognition

摘要

This paper describes an implementation of a small vocabulary isolated word speech recognition system using a recurrent neural network and some of the extensions required for a large vocabulary forms. The network operates in a self-supervised manner by adjusting an internally generated segmentation of the speech input according to the algorithm proposed by Lee et al. (see IEEE Proceedings of the International Conference ASSP, vol.5, p.3319-22, 1995) and employs the recurrent real-time learning rule described by Williams and Zipser (1989).