IF YOU have ever called a directory enquiries or flight information service, the chances are that you have spent a few happy minutes speaking to a computer. And according to some business analysts, talking to a computer in this way will soon become an everyday experience, one that changes the way we live and work.
This future relies on one enabling technology: voice-recognition software. Yet experts in the field can鈥檛 decide quite how well things are progressing. Some say it will be 50 years before a computer can truly understand what we say; others believe you鈥檒l soon be chatting to your fridge.
According to the New York-based business information company Datamonitor, the North American market for speech-recognition software will grow by more than 25 per cent each year between 2005 and 2008. Yet commercially available programs, such as IBM鈥檚 ViaVoice or ScanSoft鈥檚 Dragon NaturallySpeaking, fail to recognise a significant proportion of words. Manufacturers claim they miss around 2 per cent of all words, outside experts say it is nearer 5 per cent. In contrast, a person can expect to recognise all but 0.05 per cent of words.
Advertisement
So even under near-ideal conditions, a computer will mishear two or three words in every hundred, which at normal talking speed is about one every 10 seconds. In practice, factors such as background noise, or a speaker with a cold, can increase the error rate to 30 per cent. No wonder researchers at IBM鈥檚 T. J. Watson Research Center in Yorktown Heights, New York, have found that users of speech-recognition software spend two-thirds of their time correcting errors.
Yet improvements are hard to make. Speech recognition boils down to selecting the correct combination of words from the alternatives. So if a computer program with a vocabulary of 100 words is given a three-word phrase, it must correctly choose from the million permutations in its memory. With the average adult鈥檚 vocabulary of 50,000 words, the permutations required to recognise a few sentences becomes mind-boggling.
One short cut is to stick to a limited vocabulary, which is how today鈥檚 telephone information services work. All recognition software uses the particular waveform a spoken word creates to narrow the choice. But the final decision depends on comparing possible combinations of words in the vocabulary, however big, with how often each combination occurs in ordinary speech, and settling on the most likely phrase.
It鈥檚 a system with drawbacks. The largest phrase today鈥檚 computers can handle is three words long. So the software breaks every sentence into three-word chunks and assumes the sentence is correct if each of these make sense.
Even so, the statistical analysis is fiendish. Over the past 30 years or so speech researchers have analysed reams of text to determine the frequency of every possible three-word phrase.
As you might imagine, they have done little more than scratch the surface. The number of permutations is so huge that an embarrassingly large proportion of phrases have yet to turn up. Take three words used earlier in this article: 鈥渓argest phrase today鈥檚鈥� and type them into Google, a search engine that indexes 8 billion pages. You鈥檒l get zero hits. So a computer working out whether this was the phrase it heard would have no data to go on.
鈥淯sers of speech recognition software spend two-thirds of their time correcting errors鈥�
鈥淭his is known as the sparse data problem, and it鈥檚 the focus of a huge amount of research,鈥� says Lillian Lee, a computer scientist at Cornell University in Ithaca, New York. Lee believes the problems are so great that it will take 50 years to create a system that can match human performance.
There may be another way out. In the early 1990s, computer recognition of handwriting was just as poor. But instead of waiting for the technology to improve, the hand-held computer-maker Palm introduced a system called Graffiti, which asked the user to learn a stylised alphabet. It dramatically increased the recognition rate.
Roni Rosenfeld, a computer scientist at Carnegie Mellon University in Pittsburgh, Pennsylvania, thinks voice recognition can be improved in a similar way. 鈥淭here is no question in my mind that some type of Graffiti will emerge,鈥� he says. Rosenfeld is busy investigating how such a system might work, but it could involve using a keyword such as 鈥渁ppliance鈥� that alerts a computer to the idea that you want to switch on your TV.
If that becomes possible, then voice-recognition software could take off. We could soon be telling our fridges when to defrost the chicken and our cookers when to start the dinner. And instead of asking an automated telephone service for flight information, we could interrogate a computer over the phone about the weather in Cairo, before asking it to book a one-way flight with a window seat, and a hotel near the pyramids.
