Computers learn to talk

Press, 19 June 1979, Page 27

Computers learn to talk

From the “Economist,” London

The technology that allows computers to ape human speech is advancing so fast that a talking machine with unlimited human vocabulary may be only a few months away. It is more difficult to train the computer to hear the human voice, as well as reproduce it, but that technology is also on the move. Several new products incorporating either voice synthesis (speech) or voice 'recognition (hearing) were announced in the United States recently. One machine incorporates both. , A talking machine has to translate the waveform of the human voice into the binary digits (0 and 1) of ordinary computer talk. So the analogue signal of the human voice has to be digitised. The digital patterns of single words can thus be stored in a programmed memory’, ready for use. . .

To reproduce the original voice, when wanted, the digital signals are converted back into analogue via filters and amplifiers. A microprocessor selects the correct words — and strings them together, with appropriate pauses, to form phrases and sentences. Texas Instruments has integrated the whole system into a single, inexpensive silicon chip that can be mass produced. Last year it used the chip in a children’s toy called Speak and Spell, which can teach children to spell over 200 words. Recently Texas Instruments introduced two more talking consumer products. One is a peronal computer. The other is a hand-held language translator, a kind of pocket dictionary which helps with pronunciation problems. The options include French, German and Spanish — Chinese and Japanese are to follow.

The translator will sell for around SUS3OO, including one language module. Each module holds 1000 words, of which half can be spoken by the machine (the rest are merely displayed). Fidelity Electronics of Cichago already sells a computerised game of chess which announces the moves and captures of the game. The heart of this unit is made by Telesensory Systems Inc. (TSI), a small Palo Alto firm which specialises in making electronic aids for the handicapped. TSI designed its first speech synthesis kit fouryears ago as a talking calculator for the blind. It now sells its kit for wider uses. The Federal Aviation Administration (FAA) has bought several units for experimenting with pilot communication systems.

Many dials and displays will eventually be displaced or supplemented by speech synthesis. In industry, process control will be much easier when the operator does not have to keep looking at the display. Another application may be warning and read-out systems for aircraft and motor-vehicles. For the American driver, an end to annoying buzzers in his car; for the pilot, less need to keep monitoring instruments visually. All these systems will have one severe drawback: their vocabulary is usually limited to a few hundred words, if that. TSI is now’ developing a reading machine with unlimited vocabulary. This

will require a complex index of rules of pronunciation for the computer to refer to and a lot of new equipment. TSI expects to sell such a reading machine to the blind, for SUSIO,OOO, by the middle of next year. Several companies are tackling the problem of voice recognition. Machines that can understand the human voice have to digitise the incoming speech and match it to a pre-programmed word pattern. There are awkward variations in people’s voices: pitch, tone, etc. Regional accents — differences between male and female voices — could play havoc. Some systems under development overcome such problems by going through a get-acquainted session with each user, during which the machine stores the individual’s particular voice characteristics.

Other systems learn as they go, correcting errors by incorporating information on the new voice patterns.

Most products on the market today recognise only isolated utterances (single words or phrases). Threshold Technology of Delran, New Jersey (which is partly owned by Britain's EMI), has sold over 500 systems.

Most are used in quality control or sorting operations where the operator has his hands full inspecting an article. He can enter by voice commands like pass-fail or the destination of an article.

One system is used on an assembly line by inspectors. Their computer is located away from the factory floor and is spoken to by radio, but telephones have strange effects on voice signals, causing headaches to voice recognition computers.

No two phones will have the same effect on a voice, and each phone will vary with factors like humidity.

One company that has specialised in overcoming telephone problems is Interstate Electronics of Anaheim, California.

Recently, Dialogue Systems of Belmont, Massachusetts, launched the first system that claims to combine both speech synthesis and voice recognition in the same computer. Ore application is expected to be in banks, enabling the customer to telephone lists of accounts and amounts payable. The computer will ask, and respond to, questions put to it.

What are the chances of persuading computers to recognise a continuous stream of words, rather than a tew phrases? Some experts say that a real two-way conversation with the computer is 20 years away. But one engineer at Dialogue Systems reckons the ability to understand continuous speech could be achieved today. He says that only the high cost of the processing power holds it back. And computer people are rather good at bringing prices down.