Cross-Language Experiment
Jakub Šťastný, Pavel Sovka · Brno University of Technology Digital Library (Brno University of Technology) · 2003
Abstract. The contribution addresses the cross-language experiment. The aim was to test the possibility of the conversion French phoneme models into Czech ones. This model conversion uses the Hidden Markov Models (HMM) classification procedure. The first step consists of the iterative mapping of French models to Czech ones. The mapping is given by the analysis the confusion matrix. The second step is the Baum-Welch re-estimation resulting in the final models for Czech language. Despite of the differences between French and Czech languages the final recognition score reaches 64 % for the phoneme recognition and 74% for digit recognition. Relatively low recognition accuracy is caused by the inadequate noise model. The experiences gained with the cross-language experiment were utilized for the classification of simple human