Enrollment Training

Enrollment Training occurs when an operator does not have templates corresponding to every word in the task being run. This occurs during initial template training, any time a new vocabulary word is added to the task, and any time one or more templates have been deleted from the server. This form of training is initiated by the device.

To further reduce insertions and repeating of specific words, the number of discrete repetitions during enrollment training can be increased to improve performance with a better voice template. The more times a word is trained, the more likely the recognizer will accept it when first spoken or ignore it due to other noises. To increase the number of iterations of a word during enrollment training, either modify the Embedded Training within the task package in VoiceConsole or update VoiceApplication settings within VoiceArtisan adding an additional row for each new iteration the operator should train of that word.. In addition, for particularly problematic words, the operator may always retrain a word via the menu options to force a retrain with 10 iterations.

Update Training

Update Training is initiated with the button menu on the device. It is used to retrain a template that is not performing well. This form of training is initiated by the operator. Update training should be taught during initial voice system training and should be reinforced a week or two later.

Only the words at the current task node are available to update train. In order to update train a word, the operator must get to a point in the task where the vocabulary word is available to be recognized, and then initiate update training.

Is one training mode better than the other?

Either of these two methods can be used to train any word, but for non-digits, update training generally produces better templates than enrollment training. The following explains which method works best for digits and words:

• For digits and “ready” in the VoiceLink task, the two methods are equivalent. Both methods prompt the operator for four discrete examples of the word, and then a number of phrases containing the word.
• For words that do not contain at least six examples embedded in phrases, update training produces better results, since it will require more examples and therefore produce a better template. Update training requires the operator to say a minimum of 10 examples of each word, possibly as a combination of discrete and embedded examples. Update training is very effective at reducing insertions, because the recognizer has more examples of what variations are typical in the word and doesn’t have to be as forgiving or guess as much.