Isolated and Connected Word Recognition-Theory and Selected Applications

The art and science of speech recognition have been advanced to the state where it is now possible to communicate reliably with a computer by speaking to it in a disciplined manner using a vocabulary of moderate size. It is the purpose of this paper to outline two aspects of speech-recognition research. First, we discuss word recognition as a classical pattern-recognition problem and show how some fundamental concepts of signal processing, information theory, and computer science can he combined to give us the capability of robust recognition of isolated words and simple connected word sequences. We then describe methods whereby these principles, augmented by modern theories of formal anguage and semantic analysis, can be used to study some of the more general problems in speech recognition. It is anticipated that these methods will ultimately lead to accurate mechanical recognition of fluent speech under certain controlled conditions. I I . INTRODUCTION A LTHOUGH a great deal has been learned about the fundamental processes of speech production and speech perception, the goal of mechanical recognition of fluent speech remains elusive [ l ] -[5] . Speech recognition, however, has made major strides forward in the past decade, and it has advanced to the point where several commercial systems are currently available [6] -[ 111 . These commercial systems are predominantly isolated word, speaker-trained systems which achieve word accuracies greater than 95 percent in noisy environments. At least one system, however, is a speaker-independent, isolated word recognizer operating over dialed-up telephone lines [7] , while another is a speaker trained system that can handle a connected string of words (typically digits) [8]. In the laboratory, equally impressive advances have been recorded for speech recognition. A wide range of systems based on isolated words (both speaker-trained, and speakerindependent) have been developed for use over dialed-up telephone lines [12] -[16] ; andmorerecently speaker-independent, connected-word systems have been proposed for connected digit recognition [8] , [ 171 [ 191 . As the capabilities of the word recognizers have improved, the tasks to which they have been applied have become more sophisticated, and more difficult. Such tasks have included chess playing, data retrieval and management, airlines information and reservations, and automatic recognition of read text extracted from patents on lasers [20] -[23]. Much of this advanced research has been under the auspices of ARPA, and The authors are with BeU Laboratories, Murray Hill, NJ 07974. Manuscript received October 17, 1980; revised January 8, 1981. excellent summaries of this work are available in [4] and This paper is intended to be a tutorial in the concepts and theories underlying modern speech-recognition systems, both practical and experimental. Two aspects of the subject are given special attention. First, we treat speech recognition as a classical problem in pattern recognition and show how some fundamental ideas from signal processing, information theory, and computer science can be utilized to provide the capability of robust recognition of isolated words and simple connectedword sequences. We then describe methods whereby these principles, augmented by modern theories of formal language and semantic analysis, can be combined to study some of the more general problems of speech recognition. In particular, we show how these theories can be applied to improve the recognition accuracy of a nonideal acoustic pattern recognizer. It is anticipated that these investigations will ultimately afford accurate mechanical recognition of fluent speech provided that it is part of a “task-oriented” dialog, that is, it is restricted to pertain to a well-defined, carefully cjrcumscribed topic. The outline of this paper is as follows. We begin, in Section 11, with an overview of the pattern-recognition aspects of speech recognition. In this section, we are concerned with methods for short-time spectral estimation and elaborate on the filter bank and linear prediction methods. We also discuss other aspects of the pattern-recognition paradigm including similarity measures, temporal alignment of speech patterns, and statistical decision strategies. We then describe, within this framework, the basic word-recognition system, giving some details of its implementation, operation, and performance. Section 111 provides a discussion of the application of pattern-recognition techniques to the construction of speechrecognition systems designed to perform specific tasks. In these examples, the special requirements of each task influence the way the general theories are utilized and adapted. A relatively straightforward application of the basic principles allows us to build a voice-operated telephone repertory dialer. This particular system exploits some rudimentary task constraints by partitioning the vocabulary into sets of which only one is appropriate to any of the specific types of commands to which the system can respond. A more sophisticated kind of constraint is used to build a telephone directory listing retrieval system. Here the constraints implicit in a telephone directory are used to actually correct acoustic-recognition errors. In this context, we intro~ 2 4 1 . 0090-6778/81/0500-0621$00.75