Automatic Severity Evaluation of Articulation Disorder in Speech using Dynamic Time Warping

Abstract

Articulation disorder is a speech disorder condition where a person has an obstruction to correctly utter certain sounds. In the case of children, the early assessment and medical treatment for this disorder are most important. An Automatic Speech Recognition(ASR) based system has already been developed for assessment of articulation disorder by the Centre for Advanced Signal Processing (CASP), Department of Electronics, RIT Kottayam in collaboration with All India Institute of Speech and Hearing (AIISH), Mysore. Even though the classification of disordered speech into mild/moderate/severe categories are successfully done by the system, the numerical measure obtained for severe cases is not satisfactory. In this paper, an additional effective method to calculate a numerical measure of articulation disorder is proposed. The first method is to find out the similarity measure between articulation disordered words and their corresponding normal words by Dynamic Time Warping (DTW) of corresponding spectrogram images. Computation of Log Cepstral Distance (LCD) between disordered and normal speech is also done. Evaluation of the system is done using disordered speech in the collected dataset. It is found that spectrogram-based measure gives better results, and hence can be combined with the existing ASR-based system for better objective evaluation of articulation disorder.