Filled Pauses and Lengthenings Detection Based on the Acoustic Features for the Spontaneous Russian Speech

The spontaneous speech processing has a number of problems. Among them there are speech disfluencies. Although most of them are easily treated by speakers and usually do not cause any difficulties for understanding, for Automatic Speech Recognition (ASR) systems their appearance lead to many recognition mistakes. Our paper deals with the most frequent of them (filled pauses and sound lengthenings) basing on the analysis of their acoustical parameters. The method based on the autocorrelation function was used to detect voiced hesitation phenomena and a method of band-filtering was used to detect unvoiced hesitation phenomena. For the experiments on filled pauses and lengthenings detection an especially collected corpus of spontaneous Russian map-task and appointment-task dialogs was used. The accuracy of voiced filled pauses and lengthening detection was 80%. And accuracy of detection of unvoiced fricative lengthening was 66%.