Computers

It is relatively easy for computers to speak. A computer that says “please” and “thank you” in the right places is no miracle of science, but recognizing the words that make up normal, continuous human speech is another matter. Not until now have computers been programmed to react to a range of spoken commands. Until recently it was thought that computers would have to be programmed to the accent and speech habits of each user, and only then would be able to respond accurately to their master’s or mistress’s voice. Now rapid progress is being made with systems programmed to adapt easily to each new speaker.

The IBM Tangora system, under development at the end of the 1980s, was claimed to recognize a spoken vocabulary of 20,000 words with 95 percent accuracy. The system includes a processor that can make informed guesses as to what is a likely sentence. The system has been programmed not only with grammatical rules, but also with an analysis of a vast quantity of office correspondence. On the basis of this information, the machine can calculate the probability of one particular word following another.

Statistical probability is necessary for computers to interpret not only speech but also visual data. Security systems can distinguish between faces they have been taught to recognize, but never has a computer been able to match a human’s ability to make sense of a three-dimensional scene by identifying all objects in it.