A comparison of several variants of GMM on speech recognition task

In the automatic speech recognition task, the dominant approach is the statistical framework based on hidden Markov models in combination with Gaussian mixture models. The issues which should be solved are: how to obtain a statistically efficient estimation of model parameters, especially covariance matrix, whose number of parameters is proportional to the square of the dimensionality of the feature space, as well as sufficiently fast and accurate evaluation of observation emission probabilities. This paper provides the evaluation results of several models (diagonal approximation, maximum likelihood linear transformation, semi tied covariance) tested on Serbian SpeechDat corpus. In these experiments, the model with full covariance matrices has achieved the best performance (in terms of accuracy and computation time).