Speaker Non-speech Event Recognition with Standard Speech Datasets

Abstract

A non-speech event modelling approach to speech recognition is presented in this paper. A speaker independent spoken Czech digit recogniser is used for this purpose, and speaker generated non-speech events are modelled. Because it is important for the recogniser to be trained on suitable data, the paper shows some factors that influence the occurrence of the modeled non-speech events in the training database. Some results achieved on the analysed training database are then shown. In the experiments on forced alignment the recogniser eliminates almost all the insertion error, which is a promising property for subsequent training. However, experiments with a different basis for the non-speech event models provide almost the same results, so the difference seems to be not so significant for recognition.