Acoustic classification of Russian plain and palatalized sibilant fricatives: Spectral vs. cepstral measures

This study compares two methods for classifying voiceless sibilant fricatives forming a 4-way phonemic contrast found in Russian, but otherwise cross-linguistically rare. One method uses spectral measures, i.e. vowel formants, COG, duration and intensity of frication. The second method uses cepstral coefficients extracted from different regions inside fricatives and neighboring vowels. The corpus comprises 1,431 plain and palatalized fricatives from two places of articulation, produced by 10 speakers. Logistic regression was used to classify the productions of males and females together and separately. The productions of females yielded higher correct classification rates (highest 91.9%). Cepstral measures outperformed spectral measures across-the-board.