Classification of Pathological Speech in Speakers with Cleft Palate: Decision Tree Approach

Our work concerns the automatic classification of Pathological Speech (PP) using the Decision Trees (DT) method. The specific goal is the automatic characterization of pathological speech, particularly that produced by individuals with facial clefts, for its application in the diagnosis, evaluation, and rehabilitation of these patients in an Algerian hospital setting. We developed a corpus consisting of words containing the emphatic phoneme [ṭ] combined in 556 sound files with the Long Vowel (VL) [ā] and in 555 sound files with the Short Vowel (VC) [a]. This corpus was recorded by twenty-eight (28) speakers with cleft palate or labio-palatal cleft and a control group of thirty-eight (38) healthy speakers. Prior to this, we conducted an acoustic analysis to extract relevant features (F0, F1, F2, and F3, E0, VOT, durations [CV] and [V] of the subsequent vowel, the degree of F0 disturbance (Jitter), the degree of intensity disturbance (Shimmer), and the Harmonics to Noise Ratio (HNR)). Subsequently, a DT based on the C4.5 algorithm was applied to discriminate PP from Normal Speech (PN). The results obtained show that DTs adapt well to the recognition and classification of PP. This method allowed us to achieve appreciable Recognition Rates (RR) for PP compared to PN, with rates of 87% for VL and 85% for VC. Our work could contribute to the development of automatic diagnostics, the establishment of expert systems leading to appreciable rates of vocal anomaly identification, and support in Speech Therapy education.