Automatic Classification of Sustained Vowels Based on Signal Regularity Measures

In 1995, Ingo Titze proposed a classification scheme to classify vocalic phonemes in three types (TypeI, Type II and Type III), based on the periodicity of the voice signal. Nowadays, voices are classified by the means of spectrograms, although criteria for distinguishing among voice types are not clear yet, especially between Type I and Type II voices. Consequently, there exists great interprofessional variation in the type of voice assigned, and this also depends on each specialist’s expertise. As an approach to a more objective classification, features to discriminate between Type I and Type II voices were extracted and then used to classify voices from an annotated dataset. Classic acoustic parameters, like Jitter and Shimmer measures, and harmonics to noise ratio (HNR) were used, along with first rahmonic (R1) and two original, here proposed, parameters: principal component normalized variance (VNCP) and spectral peak-valley (PV) ratio. For the classification task, a support vector machine algorithm with linear kernel function was used, feeded with the features that minimized the cross-validation error. An error of 11.61% was obtained, with classification rates of 93.24% and 83.95% for Type I and Type II voices correspondingly.