Digital Signal Processing Reference
In-Depth Information
of discriminative power, the temporal structure of the original pattern is
preserved, too.
Discriminative training and lexicon optimization.
The simplicity of the
recognizer structure allows fast retraining of the recognition engine for a
different classification task. A set of efficient, automatic tools supports
the design of the reference knowledge under the focus of high recognition
accuracy and robustness against environmental noise.
For isolated word or command phrase recognition, an evolutionary tool is
optimizing the associative feature extraction part of the ASD classifier. The
fitness criteria, expressing the quality of a recognizer individually, are high
recognition accuracy as well as high discriminative ability between different
classes. Evolutionary algorithms outperform conventional mathematical
feature extraction methods [5,6] in that they are able to overcome local
optima and find the global optimum.
Figure 12-2.
Hands-free demonstrator on the base of verbKEYv1.6
A drawback of this method is its high resource consumption, because the
optimum often is found in several optimization attempts (5-10) after some
100 generations of 50 recognizer individuals each were evaluated. At the
