A perspective on man-machine communication by speech

A perspective is developed of the requirements for and the design of high-performance speech processing systems. These are viewed as complex systems composed of two hierarchies including a structural hierarchy of various sized units and an additional hierarchy of qualitatively distinct constraint domains. The nature of these constraints and their interaction is characterized, as well as the criteria for choice of units in the structural hierarchy. The unbounded set of surface phenomena that must be related to these units is shown to be represented by a range of patterns based on equivalence classes derived from the functional contrasts that must be maintained by the symbolic unit values. These abstract classes are related to surface correlates by complicated integrative processes, the nature of which has only recently been studied. Finally, the influence of computer science, which can be regarded as the study of complex systems, is described and the requirements for aggressive research facilities needed for further progress are developed.