September 1999, Issue 110

Taking Orders:
A Speech Recognition Module

TRAINING THE MODULE

Training begins when the Train pin is pulled to ground for at least 100 ms. When you press the momentary contact switch marked Train, you are prompted to say the first word to be trained.

Speak the word or phrase you want the circuit to recognize into the microphone. It may be up to 3.2 s long but may not contain silences longer than 0.5 s. For example, "Circuit Cellar" is acceptable as long as the two words are not separated by a long pause.

Next, the module prompts you to repeat the word or phrase. Each time a word is entered, the module creates a template. The two templates for each target word are compared and, if similar enough, are averaged together and stored in memory. If the templates are too different, an error is generated, and the module asks you to repeat the word starting with the initial template.

Before storing a word template, the new template is compared to the word templates already in memory. If the new template is too close to an existing template, the word is not accepted.

The VoiceDirect module has an automatic gain control over the audio amplifier to provide optimum signal strength. It also monitors the background noise and gives a warning if the noise is too high. A steady background noise (like a fan) has less impact on recognition than a fluctuating one (like a radio). Of course, best recognition occurs in low-noise environments.

Once a word is accepted, the module continues training by asking for more words. Training can be interrupted or stopped at any time by not speaking into the microphone at the prompt or by pressing either the Train or Recognize button.

Training is resumed by pressing the Train button. The module automatically starts training new words at the end of the previously trained words. For instance, if you trained six words and then stopped, when you resume training, the module automatically begins training at word seven.

Individual words and phrases cannot be erased or overwritten. But, the entire set of words can be deleted by simultaneously pressing the Train and Recognize buttons for at least 100 ms.