Detail of publication

Abstract

This work is concerned with the automatic speaker adaptation of an acoustic model in the automatic speech recognition system. For the model training, it is necessary to have large amount of data from many speakers. The final model, speaker independent, is then able to recognize speech from any speaker. When speaker’s identity is known, we could lower the error rate by using a model trained on the data from the particular speaker. Such a model is called speaker dependent model. The main problem with the construction of speaker dependent model is the need for large database of utterances from one speaker. This problem is often non-solvable in real conditions, however, it can be overcome by adaptation techniques. The model is adapted on the specific speaker as well as on acoustic conditions (e.g. additive noise, channel distortion) in the test utterance. The aim of this work is to discuss the methods for adaptation and the procedures of adaptation training. Some of these methods were tested and the experiments shown, that adaptation has a significant benefit for automatic speech recognition systems.