Automatic detection of vocalized hesitations in Russian speech

The article is focused on the automatic detection of the most frequent speech disfluencies in Russian speech — hesitations. Authors describe the acoustic features of Russian hesitations as well as analyze the different methods of hesitation detection. Results of acoustic analysis have shown that hesitations in Russian speech tend to be centralized, and dependent on the speech genre influence the context differently. Experiments on computerized detection of hesitations in Russian speech confirmed the efficiency and adequacy of the approaches based on acoustic information alone. Support vector machines method yielded the best results with the weighted harmonic mean of precision and recall reaching 56 %.