chapter  2
28 Pages

Phoneme Recognition Using Time-Delay Neural Networks

In recent years, the advent of new learning procedures and the availability of high speed parallel supercomputers have given rise to a renewed interest in connectionist models of intelligence [ 1 ]. Sometimes also referred to as artificial neural networks or parallel distributed processing models, these models are particularly interesting for cognitive tasks that require massive constraint satisfaction, i.e., the parallel evaluation of many clues and facts and their interpretation in the light of numerous interrelated constraints. Cognitive tasks, such as vision, speech, language processing, and motor control, are also characterized by a high degree of uncertainty and variability and it has proved difficult to achieve good performance for these tasks using standard serial programming methods. Complex networks composed of simple computing units are attractive for these tasks not only because of their "brain-like" appeal1 but because they offer ways for automatically designing systems that can make use of multiple interacting constraints. In general, such constraints are too complex to be easily programmed and require the use of automatic learning strategies. Such learning algorithms now exist (for an excellent review, see Lippman [2]) and have been demonstrated to discover interesting internal abstractions in their attempts to solve a given problem [1], [3]-[5]. Learning is most effective, however, when used in an architecture that is appropriate for the task. Indeed, applying one's prior knowledge of a task domain and its properties to the design of a suitable neural network model might well prove to be a key element in the successful development of connectionist systems.