Unsupervised learning the hidden structure of speech

Abstract

The unsupervised neural network learning procedure is applied to the analysis and recognition of speech. This procedure takes a set of input patterns and attempts to learn their function; it develops the necessary representational features during the course of learning. A series of computer simulation studies was carried out to assess the ability of these networks to label sounds accurately, to learn to recognize sounds without labels, and to learn feature representations of continuous speech. These studies demonstrate that the networks can learn to label presegmented test tokens with accuracies of up to 99%. These networks developed rich internal representation. There is no clock; the circuit is data driven, and there is no necessity for endpoint detection or segmentation of the speech signal during recognition. Training in the presence of noise provides noise immunity up to the trained level. For the speech problem studied, the circuit connection only need to be accurate to about a 3-b digitization depth for optimum performance. The algorithm used maps efficiently onto a simple VLSI hardware chip. >