The MDHMM initialization experiments (Publication 6) mentioned in the previous section pointed out that the initial ordering of the codebooks obtained by the SOM affect positively in the HMM training. In SOM the ordering of units occurs due to the adaptation of the neighbors of the BMU. During the further HMM training, the ordering will be, however, destroyed quite severely because only the best matching density components are allowed to adapt in the likelihood maximization phases (see Figure 3). If the number of mixtures is small, this closer adaptation to the training data is inevitable in order to reach the maximal modeling accuracy, since the density approximation accuracy of the Gaussian kernels decays rapidly with growing distance from the centroid of the Gaussian. Anyhow, in a large codebook several kernels can often specialize to the same input cluster and then several nearby kernels could still be adapted simultaneously to maintain the ordering. A method called the segmental SOM training is presented in Publication 5 to incorporate the use of the neighborhoods into the MDHMM training.