Admittedly, these were from the writer's (Charles J. Murray) voice, and not mine, but I gave up on IBM's Simply Speaking Gold after about a day because I just got too many ridiculous replacements similar to these.

A recent issue of Wired suggests that soon we may see online context-based searches resulting from chaining Bayesian probabilities. That means a search much more powerful than by keyword (which often sucks). We could possibly plan something like that for voice technology as well. It seems as though every time I turn around we're still five to ten years away.

Industry sometimes fails to grasp the complexity of the subject matter (e.g. the gap between computer and brain). Though each year a computer (or, an artificial intelligence program) gets closer to achieving accurate voice recognition, the gains are just as dramatic as had been forecast years ago.

It's not just computer limitations, though:

• each voice is different
• each time we say a given word, there's some variation - amplified by sickness, fatigue, anger, sarcasm or excitement
• there are gray areas in sound characteristic between certain sounds - near frequencies can represent quite different phonemes
• the complexity of the English language is quite as likely to impact voice recognition as machine translation, and for the same reasons