The Hearsay Speech Understanding System: An Example of the Recognition Process

This paper describes the structure and operation of the Hearsay speech understanding system by the use of a specific example illustrating the various stages of recognition. The system consists of a set of cooperating independent processes, each representing a source of Knowledge. The knowledge is used either to predict what may appear in a given context or to verify hypotheses resulting from a prediction. The structure of the system is illustrated by considering its Operation in a particular task situation: Voice-Chess. The representation and use of various sources of knowledge are outlined. Preliminary results of the reduction in search resulting from the use of various sources of knowledge are given. The factors influencing the structure and operation of a speech understanding system are many and complex. The report of Newell et al. (1971) discusses these issues in detail. Our own goals and efforts in this area have been described in several earlier papers (Reddy et al., 1972). The goals for our present effort were outlined in Reddy, Erman, and Neely (1970). The initial structural description of the Hearsay system was given in Reddy (1971). The model and the system that evolved after several design iterations were described in Reddy, Erman, and Neely (1972a).* The main additions to the initial proposed system were in the specification of the interactions among various sources of knowledge. In this paper, we describe the structure and operation of the Hearsay system from a different point of view, i.e., by considering a specific example to illustrate the various stages of the recognition process. Machine perception of speech differs from many other problems in artificial intelligence in that it is characterized by high data rates, large amounts of data, and the availability of many sources of knowledge. Thus, the techniques that must be * The general framework that evolved for the model is different from some previously proposed models by Liba man et al. (1962) and Halle and Stevens (1962) which imply that perception takes place through the active mediation of motor centers. Our efforts tend to support "sensory" theories advanced by Fant (1964) and others. If one modifies the "synthesis" part of analysis-by-synthesis, then our model is most similar to that of Halle and Stevens. employed differ from other problem-solving systems in which weaker and weaker methods are used to solve a problem using less and less Information about the actual task. In addition, there is …