DATA
QUALITY News....April 19, 1998

Index                     Last Week                 Contact DQ News           Latest Issue                
Featured Articles   Data & Information   Science & Technology   Education
Economics Business Law Medicine

Voice Recognition Technology: Significantly Improved

According to the April 23rd issue of The New York Times, computers can now recognize what is said to them and even talk back, thanks to relatively recent advances in speech recognition programming combined with powerful and ever-cheaper computer processing power. The two programs the Times tested were Dragon System's Naturally Speaking Preferred and IBM's Via Voice Gold.

Out of the box, the programs made mistakes; the programs are "speaker dependent." A new user has to train the programs to recognize the quirks in a voice - which typically takes about an hour. After that, the accuracy rate for both programs was about 90% at the rate of 75 words per minute. Like other current programs, these rely on a mathematical process of matching spoken words with words in a data base. Sound is first digitized. Then the pronunciation is analyzed - strings of digitized sounds are compared using a dictionary of phonetic word spellings. Finally, the voice recognition software checks to ensure the word makes sense in context. These systems can be used not only for E-mail and memos - in several Western languages - but also for editing them. Voice commands

According to the Times, speech technology is becoming useful not only for dictation and commmands, but for applications like voice-automated telephone transactions. Consumers are becoming increasingly frustrated with lengthy touch-tone phone menus. Speech recognition systems for commercial automated transaction processing (such as banking and airline reservation systems) use algorithms that work within narrow contexts, restricting the size of vocabulary and extracting the core meaning. These systems use many tactics including word-spotting algorithms that can handle variations like "this Tuesday" and "next Tuesday."

Although the IBM and Dragon speech recognition systems are relatively inexpensive ($149 and $220 respectively) they will only work with Windows 98/95/NT. And they require a Pentium-based personal computer of recent vintage, a sound card, and large amounts of RAM and free disk space. The article was written by Anne Eisenberg and appears on page E1.