Dictating the future of personal computers: Improved voice recognition allows hands-off word processing

CAN'T TYPE? Tired of correcting spelling mistakes? Like to talk as you work? Help is now at hand, for those who don't mind speaking slowly. Last week IBM began shipping the Personal Dictation System, which allows you to enter text into a personal computer by speaking. Users dictate into either a hand-held or headset microphone, and the screen displays their words as they talk. The text can then be transferred into a number of standard word-processing packages.

The Personal Dictation System needs to be hooked up to at least a 486 personal computer. The computer also needs to be fitted with a Dictation Adapter, a speech card that converts analogue signals from the microphone into digital code. The basic system has a vocabulary of 32,000 words, with additional technical vocabularies available for journalists and various medical practitioners. Further vocabularies for lawyers and doctors are under development.

At present the system is on sale only in the US, at a cost of about dollars 1,000, but will be available in Europe later this year in UK English, Spanish, German, French and Italian versions.

Before the dictation system will work, each user has to train the computer to understand his or her voice by reading to it for 90 minutes. The program then builds a mathematical model of the individual's voice pattern to take account of accent and speech characteristics. When the user dictates into the machine, the speech waveform is digitised and matched with a library of word models.

This pattern-matching approach was rejected by early researchers into speech input systems in the 1960s in favour of rule-based artificial intelligence systems, because it requires huge computing power. Rapid increases in computer technology mean this power is now available on the desktop.

The system can cope with no more than 70 words a minute. Each word must be distinct, with a pause between each. Talking this way is an acquired skill and seems tortuously slow. However, non-professional typists rarely type accurately at this rate, and once the words are accepted there should be no spelling errors.

Other features include a Voice Action Editor, which enables users to create personal instructions. For example, a lawyer could produce a standard disclaimer. Whenever that paragraph has to be inserted in a letter, 'standard disclaimer' are the only words needed. The system can also be taught commands such as 'bold type' or 'new paragraph'. It understands all the commands in the computer's menu.

Speech input has been a long-term goal of computer scientists, but so far systems have been too slow and error-prone to gain widespread commercial acceptance, and have mainly been used by disabled people who could not type at all. Many other computer companies remain doubtful that speech recognition can be made accurate enough for widespread use. They also argue that the latest computer interfaces make machines so easy to use that speech input is irrelevant.

But IBM says the Personal Dictation System will be particularly useful for those who need to use their hands while working. They will be able to dictate instructions or reports at the same time. For example, a radiologist could report on a series of X-rays by speaking into the headset microphone while examining the film. The system could also be used by people who have suffered repetitive strain injuries using computer keyboards.

IBM's confidence in the Personal Dictation System is partly based on a similar system that has been available on its workstation computers for the past year. The company says that more than 70 software companies are committed to developing applications based on its speech technology.

A series of speech recognition products will be launched in the next few months. Elton Sherwin, the market development manager for speech recognition at IBM, says: 'What we can do today is already radically different from what we could do even two years ago.

'We only became comfortable with accuracy for double numbers like dollars 14.40, dollars 15.50 in July 1993. But perfecting the recognition of numbers will allow sophisticated financial management by phone or cable.' As a result, he believes, speech recognition systems will be available on interactive cable systems in the US within 18 months - for paying bills, playing games, ordering movies and so on.

Although these applications will require speaker independence - operating without first being trained to understand each voice, and with continuous speech capabilities - the vocabulary required will be very limited. IBM already has a continuous speech toolkit for developing applications which can be used with a 1,000-word active vocabulary chosen from a base of 20,000 words, and the next stage will be to adapt this for the commands needed to operate a cable TV system.

IBM is also testing the continuous speech system in collaboration with police forces, so that, for example, police officers can ask the computer in their car to search for a registration number while chasing a suspect, or request background information about someone they have detained. Other software companies are using the system to develop applications for casinos, court reporting, health care and financial services.