Dysarthric speech detection from telephone quality speech using epoch-based pitch perturbation features

Dysarthria is a motor speech impairment that impacts verbal articulation and co-ordination. Detecting dysarthria is a primary and essential step for early diagnosis and treatment. In this paper, we attempt dysarthric speech detection from telephone quality speech by using pitch perturbation (PP) measures computed with the recently introduced continuous wavelet transform (CWT)-based epoch extraction approach. This approach has the strong advantage that it is highly robust to telephone channel degradations. Six PP measures were computed from the extracted epochs. For comparison, the PP measures were also derived using two well-known epoch extraction methods, namely, zero-frequency filtering (ZFF) and dynamic programming phase slope algorithm (DYPSA). The experiments were carried out using the TORGO dysarthric speech database, which consists of speech from 7 healthy speakers and 8 dysarthric speakers. The G.191 software tools were used to convert clean speech to telephone speech. The results show that the PP measures computed with the CWT-based approach can better discriminate dysarthric and healthy speakers under telephone environment than those extracted with the other two epoch extraction methods.