You are here:

Voice Recognition Technology Boosted

Editor's Note: The following article is reprinted from the Toronto Star, November 26, 2004, courtesy of Torstar Syndication Services.

ViaScribe, the latest voice recognition software from IBM, could help level the playing field for students with differing abilities when it becomes available. Since the mid-'90s, VR technology has held the promise of making a big difference in the lives of disabled people. But the software hasn't come close enough to replicating the complex hearing ability of humans. It misses or mistakes words, fails to pick up accents and requires very slow speech. VR software acts like the human ear, listening to the spoken word and then automatically converting it into text so the hearing impaired can read on a monitor what has been said. For the visually impaired, ViaScribe (which has not been released yet) takes standard VR technology even further, allowing blind students to automatically record lectures into audio files that can be played back whenever notes have to be reviewed. The software was demonstrated last week in Toronto at IBM's presentation facility on Bloor St. E. Professors who tested ViaScribe were getting well into the 90 percent range as far as accuracy, says Sara Basson, a program manager with IBM. "We're making great leaps as far as accuracy. In museums (where the software was tested during tours, using hand-held text and audio devices for the hearing- and visually impaired) and other settings we're getting close to 100 percent accuracy." There are many bottlenecks in trying to perfect VR, says Parham Aarabi, a sound and speech-processing specialist with the University of Toronto's computer engineering department. "The main one is processing power." He says the calculations needed to replicate intricacies the human ear uses to filter and recognize sound is still beyond the capacity of VR technology. Another obstacle is the ability to create software that accounts for the environments in which speech is expressed. "If you're in a lab, with a controlled environment, you can achieve 90 percent accuracy rates. There are many manufacturers of the software that say their product has 90 to 95 percent accuracy." However, in natural, uncontrolled settings, he says the rates are more like 60 to 70 percent. Experts set a standard between 85 and 90 percent accuracy for voice recognition technology to be useful.