When carefully trained by the maximum likelihood methods or using the SOMs, the resulting HMMs can be quite accurate in modeling the characteristics of the phonemes. However, the models are not optimized to discriminate between the states of different models. Some phonemes might resemble each other very much in certain contexts and thus the corresponding models may respond equally well. The selection of the highest likelihood model can then be based on some random details giving arbitrary decisions.
The LVQ was studied in order to develop a training method which would learn the essential information needed to discriminate the phonemes from the training samples in a simple adaptive manner. The method should tune the HMM parameters so that the correct models would stand out with as wide marginal as possible to avoid the risk of sequence misclassifications. Anyhow, any side effects like instable learning behavior or loss of generalization should be avoided. A possible loss of accuracy in the probability values for all competing sequences is not of concern, because the objective of this ASR subtask is only to find out the correct phonetic transcriptions regardless of the correctness of the probabilities.