Dragon roars, IBM much improved
- — 11 January, 2000 16:17
The newest releases from Dragon Systems and IBM, two key developers of voice-recognition software, continue this forward momentum with significant updates to their flagship products. We examined Dragon Systems' $349 Naturally-Speaking Preferred 4.0 and IBM's $369 ViaVoice Pro Millennium. Dragon's effort continues to remain the overall accuracy champ.
Both new products require a 333MHz Pentium II with 64MB of RAM as a realistic minimum; a faster CPU and 128MB of RAM are even better. You'll need substantial disk space, too. NaturallySpeaking occupies 171MB, ViaVoice 310MB.
To determine whether the faster PIII and its SIMD (single-instruction multiple-data) extensions improved performance, we tested the packages on a 333MHz Pentium II and again on a 550MHz Pentium III, both equipped with 128MB of RAM. (The Dragon package has special SIMD support; neither product uses the 3DNow instructions included in AMD's Athlon processor.) Both packages displayed dictated words a bit faster on the PIII system than on the PII, but the PII was very usable. More important, the change in CPU didn't affect accuracy or training time.
Over the past year, voice recognition software companies have slashed the amount of coaching their software requires. Earlier, setup meant reading to your PC for 30 minutes or more to teach it to recognise your voice.
These new versions shorten training time markedly: in our tests, NaturallySpeaking 4.0 required about five minutes of training, and ViaVoice about eight minutes. Your results may vary.
But accuracy is the crucial test in voice recognition. And while the vendors claim small improvements in these new versions, the programs continue to miss a lot of words. So if you're dictating a long document, you can expect to spend some time correcting errors.
Correcting typos by voice remains an ordeal. Moving a cursor by voice command is frustratingly time consuming, and a small vocabulary of prescribed editing jargon needs to be mastered ("scratch that", "correct that", and so on). Voice control of menus and dialogue boxes is still cumbersome as well. Barking commands like "click file" and "click save" or intoning a file name can be an awkward way to interact with your PC.
Both products let you use voice commands to browse the Web. To navigate in NaturallySpeaking, you say the name of the link. ViaVoice assigns a number to each link on the Web page you're visiting, and you activate the link by saying the number. We prefer Dragon's approach because the IBM method can clutter Web pages with the superimposed numbers. We found both packages' Web-browsing features quite useful for activities like communicating in chat rooms.
In dictation tests, Dragon NaturallySpeaking was the clear winner, consistently topping 95 per cent on accuracy - which means it missed about 1 in 20 words.
Besides defending its position as the most accurate voice-recognition product, Dragon made several functional improvements in version 4. Its predecessor, version 3.5, integrated with only a few programs, while version 4 supports most programs except design and drawing packages.
In accuracy, IBM continues to play catch-up with NaturallySpeaking, achieving a respectable accuracy rate of just over 90 per cent - missing about one word in 10.
ViaVoice Pro Millennium Edition bears an improved interface. In particular, we liked the Recognition Wizard. This automatic help tool pops up whenever ViaVoice detects problems, such as repeated misrecognition of the same words. Previous versions of ViaVoice integrated with a wide variety of other applications but often had problems distinguishing between commands and dictated text. This particular snag occurs far less frequently in the latest version.
NaturallySpeaking Preferred Version 4
Phone: (02) 9238 6500
Distributor: Marketing Results
Phone: (02) 9899 5888
Distributor: Voice Perfect Systems
Phone: 1800 063 243
IBM ViaVoice Millennium Edition
Phone: 13 2426