Stop consonant voicing classification for computer-assisted speech training of patients with cleft lips and palates

Cleft lips and palates (CLP) may cause functional disorders even after adequate surgical treatments, speech disorders being one of them. Automatic algorithms utilizing acoustic-phonetic knowledge are needed in developing computer-based tools for assisting the speech training of CLP patients. This work focuses on acoustic discrimination among voiced, voiceless unaspirated, and voiceless aspirated stop consonants with the same place of articulation and aims at revealing a set of acoustic measurements capable of discriminating a CLP patients' speech. Acoustic measurements based on duration and signal energy are proposed and studied. Analysis of variance and classification experiments demonstrate high potentials in using these acoustic measurements in developing automatic voicing classification algorithms for speech training tools. The overall classification accuracy of 92% is achieved in classifying non-CLP data, in which the best result obtained is 99% for the alveolar case. The proposed measurements can classify data from CLP patients with almost 90% accuracy even when the classifier is trained only on the non-CLP data.