A discriminant measure for model complexity adaptation

1 ABSTRACT We present a discriminant measure that can be used to determine the model complexity in a speech recognition system. In the speech recogition process, given a test feature vector the conditional probability of the feature vector has to be obtained for several al-lophone (sub-phonetic units) classes using a gaussian-mixture density model for each class. The gaussian-mixture models are constructed from the training data belonging to the allophone classes, and the number of mixture components that are required to adequately model the pdf of each class is determined by using some simple rule of thumb { for instance the number of components has to be suucient to model the data reasonably well but not so many as to overmodel the data. A typical example of the choice of the number is to make it proportional to the number of data samples. However, such methods may result in models that are sub-optimal as far as classiication accuracy is concerned. In this paper we present a new discriminant measure that can be used to determine in an objective fashion, the number of gaussians required to best model the pdf of an allophone class. We also present the results of experiments showing the improvement in recogntion performance when the number of mixture components is chosen based on the discriminant measure as opposed to the rule of thumb. These results are presented both for the speaker-independent and speaker-adapted case. 2 INTRODUCTION We present a discriminant measure that can be used to determine the model complexity in a speech recognition system. In the speech recognition problem, feature vectors are extracted periodically from the input speech and are matched to diierent sequences of phones, that represent words in the vocabulary. In the statistical approach to speech recognition, this is done by estimating the probability density of each phone in the feature space from the training data, and using these pdf's to assign a probability to a test feature vector. The most common case is where a parametric model is used to model the pdf, with the parametric model generally being a mixture of gaussian distributions. Hence, in the speech recogition process, given a test feature vector the conditional probability of the feature vector has to be obtained for several allophone (sub-phonetic units) classes using the gaussian-mixture density model for each class. It is not unusual to use tens or even hundreds of thousands of diierent …