IBM Corp claims to have improved the accuracy of its ViaVoice range of desktop speak recognition software with changes to the way the software ‘learns’ voice models and improved microphone and audio specifications. The company has also added voice- enabled web, email and chat features to the versions of the software.
The company has three new versions of the desktop software – which it has dubbed the Millennium line – common to all are the improved accuracy features. Chief among these is continuous analysis of a users’ speech as they use the program. A user’s underlying voice model is regularly updated using this information.
IBM has also tried to improve the collection of voice data at source, by including a new microphone with better signal-to-noise ratios in the packages. We’ve done a lot of work on acoustic data, said Helene Linden, product line manager for IBM Speech Systems. The package includes audio level detection, which allows the user to optimize the levels for the best voice recognition performance.
Linden wouldn’t give any actual statistical figures for the claimed improvements in accuracy. She bemoaned the lack of standardized tests for speech recognition products. Nor would she be drawn on what IBM’s share of the market was against rivals such as Lernout & Hauspie Speech Products NV or Royal Philips Electronics NV.
However, IBM is well placed to move its ViaVoice off the desktop and into mobile devices through the work of its Pervasive Computing unit. Linden said that she expected keyboardless mobile devices to be one of the market drivers in the future of voice recognition. All the major manufacturers of desktop software now claim 90% plus recognition rates for their software and there is some question as to how much more performance can be squeezed out of software alone. Linden said that she thought that working on microphone technologies and the raw speech engines that sit behind all voice-enabled products could wring more improvements from desktop packages.
The minimum system requirements for the ViaVoice software are a 233MHz Pentium processor with 256k level 2 cache, a 16-bit soundcard and a 4x CD-Rom drive. Software running on Windows needs a minimum of 48Mb, 64Mb on NT. As announced at MacWorld in July, IBM will also introduce an entry-level package for the Mac. The standard Millennium package costs $60, the web-enabled edition $80, while the Pro version that brings together of the features retails at $179. The ‘American English’ version will be available next week, with European language versions rolling out in October.
