Optimal acoustic model complexity selection in polish medical speech recognition

In the paper, the method of acoustic model complexity level selection for automatic speech recognition is proposed. Selection of the appropriate model complexity affects significantly the accuracy of speech recognition. For this reason the selection of the appropriate complexity level is crucial for practical speech recognition applications, where end user effort related to the implementation of speech recognition system is important. We investigated the correlation between speech recognition accuracy and two popular information criteria used in statistical model evaluation: Bayesian Information Criterion and Akaike Information Criterion computed for applied acoustic models. Experiments carried out for language models related to general medicine texts and radiology diagnostic reporting in CT and MR showed strong correlation of speech recognition accuracy and BIC criterion. Using this dependency, the procedure of Gaussian mixture count selection for acoustic model was proposed. Application of this procedure makes it possible to create the acoustic model maximizing the speech recognition accuracy without additional computational costs related to alternative cross-validation approach and without reduction of training set size, which is unavoidable in the case of cross-validation approach.