February 1998, Issue 91

Low-Cost Voice Recognition

TINY APPLICATIONS

For testing purposes, the system was trained with eight words: "VCR", "television", "telephone", "stereo", "CD", "PC", "yes", and "no". Each word was trained twice, thereby occupying 16 templates.

Recognition accuracy approaches 100% when background noise isn’t too severe. It also works with ~90% accuracy using speakers who didn’t train the system.

A speaker-independent vocabulary can be constructed by having multiple trainings of a few words. For example, training "yes" and "no" eight times over a set of different speakers yields excellent results.

A note of caution: when using Tiny Voice, don’t use a lot of short words (e.g., the numbers "one", "two", etc.). They’re a bit beyond its capabilities.

And watch for commands that sound alike. For example, "on" and "off" will get you in trouble. Instead, try "turn on" and "off please".

A fun application might be a voice-activated padlock. Change the code so you have to enter one, two, or three voice commands in sequence. Then, multiply the scores. If the result is small enough, then "open sez me."