Tolerance for data defects.

The smoothness of an output density of an HMM state is important for several reasons. The training data is finite and even with large and carefully balanced databases, there will almost always be input space areas with insufficient presentation. Training the codebook first using a large neighborhood exploits in a way some additional input samples for each unit. In addition to providing less random parameter values outside the main input data clusters, this may also allow to train larger codebooks by the same available training material, because the same samples could be used in training of larger amount of units.

In K-means based training it may sometimes happen that part of the available modeling capability is left unreachable, because some of the Gaussian kernels will be practically unused. In SOM, however, the first training epochs with large neighborhoods draw automatically most units near the important areas and enable thus their later use. The advantage of the maximal exploitation of the modeling capacity is revealed, when the training reaches the stage of closer adaptation. As the training proceeds the quality of the result does not improve like it should, if the initialization has been insufficient (e.g. see Publication 6).