Sixth International Conference on Spoken Language Processing
This paper describes the combination of the work of speech therapists and speech recognition systems. Our long term goal is to evaluate the degree of stuttering during therapy and to use the automatic analysis of stuttered speech as a screening method, e.g. the search for potential stutterers at an early age. The approach is to have a patient read a standard text aloud and then automatically count the unfluent parts and classify them. The text to be read by the patients is automatically transformed into a formal grammar that considers potential dysfluencies caused by stuttering.
Recordings from stutterers were compared to recordings of nonstutterers. Word and phoneme accuracies of the stuttered text in relation to the number of detected dysfluencies showed correlation coecients of up to 0.99. Recordings from stutterers contained much more pauses in a wider time range than from nonstutterers, especially in the interval up to 200 milliseconds (factor 10), and between 200 and 500 milliseconds (factor 2). The sum of the durations of all detected pauses and the number of repetitions were set into relation. The results seem reasonable for a distinction between stutterers with many repetitions/short pauses and stutterers with few repetitions/long pauses.
Bibliographic reference. Nöth, Elmar / Niemann, Heinrich / Haderlein, Tino / Decher, M. / Eysholdt, Uwe / Rosanowski, F. / Wittenberg, T. (2000): "Automatic stuttering recognition using hidden Markov models", In ICSLP-2000, vol.4, 65-68.