September 1999, Issue 110

Taking Orders:
A Speech Recognition Module

SPEAKER DEPENDENCE

Speech recognition is classified into two processing categories—speaker dependent and speaker independent. Speaker-dependent speech-recognition systems are trained by the person who will be using the system. These systems achieve a high command count and better than 99% accuracy for word recognition.

One drawback, however, is that the system responds accurately only to the individual who trained the system. But, an important advantage is that the circuit may be trained in any language.

Actually, language isn’t even necessary. A series of grunts and whistles (as long as they can be repeated accurately) can be used in place of words. This is helpful to people who, through accident or illness, have lost the ability to verbalize words.

The VoiceDirect module is speaker-dependent, which is the most common approach employed in software for PCs. Sensory also offers other chips for use in speaker-independent modes.

A speaker-independent system is trained to respond to a word regardless of the speaker. This system must respond accurately to a large variety of speech patterns, inflections, and enunciations of each command word.

The command-word count is typically much lower than the speaker-dependent systems, but high accuracy can be maintained when system demands are constrained by a limited number of commands. Industrial applications more often require speaker-independent voice recognition systems, such as the systems used by AT&T and other telephone companies.