Enter the Dragon

My trip to China had the unexpected effect of forcing me to reappraise the human-computer interface in general, and voice recognition software and the Tablet PC in particular. The rush to get my work done produced ominous twinges of the RSI (repetitive strain injury) which I suffered a few years ago. And it happened that Scansoft, which bought assets of the troubled voice recognition specialist Lernout and Hauspie late last year, had just sent me the latest version of its dictation software Dragon Naturally Speaking. I resolved to make a serious attempt to see if it could take some of the strain out of my arms and wrists.

Prototype Tablet PC design by Via Technology

I have never succeeded in getting any useful results out of dictation software, though I know people who have. It can work for anyone prepared to persevere, particularly those with disabilities that preclude them from typing. Lawyers, doctors and others who do a lot of dictation also seem to benefit - perhaps because they use a relatively small, albeit highly specialised, vocabulary for which they can purchase tailored products.

But dictation software takes almost as much learning as the keyboard, and is not as versatile, so most people stick to typing. The new Dragon software is supposed to embody the best of both L&H's Voice Xpress and its bought-out long-time rival Naturally Speaking, with a few added tweaks. PCs are also a lot more powerful than when I tried an earlier version.

Still it was more in hope than expectation that I loaded the software onto my home PC, powered by a 1.4GHz Athlon, with 512MB of Ram. To my astonishment, after just five minutes of training it transcribed the first few sentences I spoke nearly word perfect. On previous tests I have generated little but gibberish. But my initial enthusiasm quickly faded when I realised the limitations.

Headsets need to get mobile too The placing of a good microphone near the mouth is essential if you want to use dictation software. But this can become uncomfortable, particularly as all the headsets sold with dictation packages tie you to your PC - there are as yet no wireless headsets.   Those you can buy separately, such as the Bluetooth-enabled Plantronics M1500 pictured above, are designed for mobile phones - you can't plug their remote modules into a PC. Incidentally the M1500 is due for release this summer for about £140 - almost as much as the software. Research is going into directional mics that can sit on the table, but they have yet to make it big.   Dragon will transcribe dictation from many makes of voice recorder or minidisc players, though the results are said to be less reliable than direct dictation. Alternatively it will transcribe from a .WAV file - just don't count on the results.

You need to place the microphone precisely to one side of the mouth (right), and extraneous noise can produce errors. Dictation would also clearly be impracticable in a crowded, open-plan office.

More serious is the fact that the software, which relies heavily on the statistics of speech, has no sense. It recognises and uses context but does not understand it. The result is that it transcribes whole sentences better than individual words or phrases. It tries to be helpful, offering alternatives when it gets something wrong, but correcting mistakes or making alterations can be very frustrating indeed. This virtually rules out the kind of cut-and-paste juggling that makes the word processor such a powerful tool for writers. However, whether because of better algorithms or better machines, Dragon dictation software seems a lot more useful than it was.

Speech control of the computer is more viable and you can get round the limitations of dictation by writing and editing your stuff in longhand and then dictating a fair copy to the computer - at least it saves your wrists.

Dictation software is unlikely to get much better soon, because we are far from being able to give the computer the cognitive skills human stenographers bring to transcription - and even they get it wrong sometimes. But the fact that the software is beginning to work at all is exciting.

The current human-computer interface is primitive, at the start of what will clearly be a very long evolution to something at least as multi-modal and subtle as human-to-human communication. I do not expect much initially of Microsoft's Tablet PC, which was being previewed as we went to press, but I believe it could precipitate the next phase of interface evolution by promoting a more usable pen-driven platform than the underpowered, undersized handhelds of today. It will also have some native speech capability.

Traditional handwriting recognition (as opposed to a purpose-designed, unambiguous computer script) will hit the same impasse as dictation software, for the same lack of human-like cognition. But a combination of speech and handwriting recognition, the pen's ability to point and select, and hardware powerful enough to run it all simultaneously, could add up to something much more than the sum of its parts.

A tablet driven simultaneously by pen and voice would start to use all that processing power Intel and AMD have been pushing at us - and it would require all the memory it could get. But if a tablet is to be carried about it will also need to be able to run full power without burning flesh.

First-generation tablets are unlikely to be able to offer such energy-efficient computing power, and when it comes it may entail something very different from the traditional CPU to process the speed and handwriting.

The good news for the chipmakers is that pen and speech input could end the short era in which hardware has run ahead of the needs of most software, and make upgrades a necessity rather than a luxury.

* Dragon Naturally Speaking version 6.0 costs £147 inc VAT with a headset. Watch out for more on the Tablet PC next month.(In the printed edition of this story we gave this price for the Professional version, which costs £647 inc VAT and includes many cusomisation features. The recognition engines are the same.)

Tablet handwriting recognition disappoints