Pattern search in dysfluent speech

Pattern recognition in time series is often used in data mining and in bioinformatics. Speech can be considered only as a different type of signal and processed as a time series. Stuttered speech is rich in events also known as dysfluencies, typically repetitions. This paper describes a new method for enumerating complex repetitions. Classical approaches to stuttered speech analyzed dysfluencies in very short intervals, which were sufficient for recognizing simple repetitions of phonemes. However, the problem of repetitions of syllables or words was typically ignored due to high computational demands of classical methods for analysis of longer intervals. Our approach uses a method adopted from data mining and bioinformatics, together with efficient representation of speech signal, which simplifies processing of speech enough to enable analysis of longer intervals. Results show applicability of the proposed method.