Detail of publication

Abstract

This work is focused on the automatic speaker adaptation of an acoustic model, which is a part of the automatic speech recognition system. To train the acoustic model it is necessary to have large amount of data from many speakers. The final speaker-independent model is then able to recognize the speech from any speaker. The speaker-independent model is adapted to the speech of a specific speaker. Ordinary adaptation techniques introduced in this work perform poorly in cases with insufficient amount of adaptation data. The aim of this work is to discuss methods of adaptation and adaptation training. To avoid the problem with lack of adaptation data various robust solutions have been described and new one have been proposed. Some of these methods were tested, and the experiments show that the robust adaptation contributes significantly to the task of automatic speech recognition.