BANDWIDTH EXTENSION OF TELEPHONY SPEECH - Adaptive Signal Processing: Next Generation Solutions - page 374

Figure 7.13 Structure of the vocal tract transfer function estimation using a neural
network.
from the bandlimited to the broadband envelope. Also the temporal transition model
between adjacent frames is learned.
The training of the neural networks is accomplished similar to the computation of
the mapping matrices of the last section by providing pairs of bandlimited and broad-
band data. The training itself then has to be performed very carefully to avoid over-
training. Overtraining denotes the optimization on the training set only without
further generalization. This can be observed by using a validation data set to control
if the network still is generalizing its task or beginning to learn the training data set
by heart. Figure 7.14 shows such a typical behavior.
The optimum iteration to stop the training is marked in Figure 7.14. This optimum
is characterized by the minimum overall distance between the actual and the desired
output of the artificial neural network produced using the validation data set as input.