Abstract
Even with over thirty years of research, the goal of being able to speak directly to a computer in a completely natural manner has proved to be singularly elusive. Consequently, a wide range of techniques have been tried in order to avoid or overcome the highly complex problems involved. It is possible however, to broadly classify all the machines that have been developed into one of three categories, where the categories are defined by the type of spoken material the machines are designed to accept. Isolated Word Recognisers require the user to segment his speech, thereby avoiding many of the problems of recognising continuous speech. A few hundred such machines are in use world-wide, mainly in specialised data-entry or control applications. The advent of one or two chip systems will increase the application potential for this kind of machine. Connected Speech Recognisers allow the user to speak in a more natural manner and enable faster data input, but still do not permit the transcription of a wide range of normal human speech. This has to be done by Speech Understanding Systems, which so far exist only in research laboratories and fall a long way short of human performance in terms of vocabulary size, error rate and speed of operation. This paper will outline' the problems and status of each of these three types of machine, and discuss the possibilities for the input of data to computers by means of speech.