X
Tech

The Gadget Showdown: 'Cocaine, Mrs Phillips?'

Speech recognition hasn't yet matured as an everyday technology, but its use is growing. Ben King weighs up the two most popular desktop packages...
Written by Ben King, Contributor

Speech recognition hasn't yet matured as an everyday technology, but its use is growing. Ben King weighs up the two most popular desktop packages...

"OK Mrs Phillips, take dictation..." It's not a sound heard very often in the modern office, least of all in the high-tech sector - the sound of a busy executive angrily dictating a business letter to a terrified secretary. Most people find it easier to shout at each other over the phone or converse in a string of badly spelled emails. Still, technology takes away with one hand, and gives with the other - it's becoming increasingly popular to dictate direct to a PC, using a speech recognition program. silicon.com decided to check systems currently on the market. The two leading products are NaturallySpeaking 5 from Dragon Systems, now a subsidiary of troubled Belgian company, Lernout and Hauspie, and IBM's ViaVoice, now at release 8. These offerings are surprisingly similar. Both boxes contain a CD with the software and a headphone set with a boom mike. The first stage is to "train" the software to recognise your voice. For some reason, both programs require the reading of a passage from Alice In Wonderland, but Dragon gives a couple of other choices, including a rather depressing selection from Charlie and the Chocolate Factory. It took two hours to train IBM, during which time I became so angry with its failure to recognise simple words I started shouting at it and had to have a cup of tea to calm down. Dragon only makes you talk for about 10 minutes before it starts recognising text. Both have an option to do more training for greater accuracy. Once trained, both programs work in much the same way. There is a little bar at the top of the screen, above whatever application that's running. You can then open a word processing program, click the "Start Talking" button, and get busy. I tested the program with two different passages from a popular current affairs magazine. Out of 200 words, IBM produced 37 errors and Dragon produced 26 - a surprising result, as I was expecting IBM to perform better, having spent so much time and energy training it. Yet on balance, IBM is the marginally better product, despite an exceedingly annoying little pencil that pops up and gives advice, knocking the Microsoft paperclip into the shade with its irritation power. Other than that, it seemed slightly faster, better designed and easier to use, despite Dragon's better score in the dictation test. I'm not a particularly fast typist but both speech programs were significantly slower than writing by hand, once the time taken to correct all the mistakes is included - about five times as long, in fact. So most of this article was written with fingers, not vocal cords. On the other hand, it is quite nice to be able to sit back and stare out of the window, just letting ideas flow out over the page, rather than staring at a cathode ray screen. If the headphone lead is long enough you could even refine your golf swing as you write a letter, just like company bosses did in the good old days. That doesn't mean speech recognition will only appeal to compulsive golfers and people who have recently sustained serious arm injuries. Like typing, SR takes a lot of learning. Users must train themselves as well as their wares, and in a few weeks the software starts responding well. You can also spend a bit of extra time training the computer - reading more Alice In Wonderland - to let it get to know your voice better. Here are a few tips to help the system work better: - Speak in the most boring voice you can: constant speed, constant emphasis, constant enunciation. Try to emulate John Major and you won't go far wrong. - Don't try and dictate word by word. The computer uses the context to help make sense of what you say, so if you speak at normal speed it functions much better. - Use a powerful PC. These tests were done on a PC with a 533MHz processor and 64MB of RAM. This set-up was able to deal with both programs quite comfortably, but you wouldn't want anything slower. - A big hard drive is also useful. SR files take up a phenomenal amount of space. The Dragon files on the hard disk were 135MB, and IBM's a whopping 313MB, which will grow the more training I do. The whole of MS Office only takes up 38MB. If conditions aren't absolutely quiet, SR will really struggle. The tests were done at home, not in an office. Finally, proof read very carefully. The SR can produce some truly bizarre errors. "OK, Mrs Phillips" came out as "Cocaine, Mrs Phillips" the first time round. Could this be the first time a computer has made a Freudian slip?
Editorial standards