Rev Your Speech Engines With IBM and L&H

How times have changed in the world of speech recognition software. Gone are the days of tedious, hour-long "training" sessions in which you read passages aloud so that the software could get acquainted with your voice, all in the interest of improving accuracy.

New releases from IBM and Lernout & Hauspie whittle down that convoluted process considerably. Both IBM's $200 ViaVoice for Windows Pro Edition and Lernout & Hauspie's $199 Dragon NaturallySpeaking Preferred 5.0 can have you up and running in about 5 minutes.

After running both packages through marathon talking tests, I determined that NaturallySpeaking delivers better word recognition, processes dictation faster, and is generally easier to navigate. However, ViaVoice offers some unique features that NaturallySpeaking lacks.

Tools for Talking

IBM's recommended hardware requirements are a tad stiffer than L&H's. To run ViaVoice on any version of Windows, you need at least a Pentium II-300 or an AMD-K6; to use NaturallySpeaking, you need a Pentium II-266 or an AMD K6-2. ViaVoice gobbles up 510MB on your hard disk and requires 48MB of memory with Windows 95/98/Me, 64MB with NT 4.0 (with Service Pack 5), or 96MB with Windows 2000. NaturallySpeaking needs 150MB of disk space and 64MB of memory, no matter what flavor of Windows you have. Both products require a standard 16-bit sound card with a microphone input jack, or a Universal Serial Bus port to use with a USB microphone.

Each product includes high-quality noise-canceling headsets. IBM's offering, the Andrea Anti-Noise NC-61, is far superior, though, as it has a comfy, sturdy headband along with a cushy earpiece. In addition, Andrea's headset attaches to a USB adapter and sound kit box that lets you circumvent your computer's sound card inputs--a bonus for users who dictate to notebooks, which are notorious for having poor sound input/output quality. L&H ships a rather flimsy Telex headset, which uses an unpadded metal strip as the headband.

Both products offer easy-to-follow audio setup wizards that work better than their predecessors did. ViaVoice shows you a particularly helpful 3D image of how to position your microphone headset. L&H gives you more entertaining training excerpts to read from, such as selections from Charlie and the Chocolate Factory and Dogbert's Top Secret Management Handbook, versus IBM's Treasure Island and Growing Up Digital.

I tested the software by dictating a series of e-mails, business reports, and newspaper articles. I ran the tests on a Pentium III-600 notebook with 256MB of RAM, a Pentium II-400 desktop with 64MB of RAM, and a Pentium II-333 desktop with 96MB of RAM. Not surprisingly, the PIII-600 speedily processed my verbal ramblings into text.

Although NaturallySpeaking and ViaVoice now have accelerated training (or "enrollment") requirements, the more you use the products and retrain them as you go along, the better each gets at recognizing the nuances of your voice. Both programs improved over time as a result of additional enrollment and ongoing correction of misrecognized words on-the-fly.