Spectral approximations for sibilant classification

This paper deals with an analysis of power spectra of prolonged Czech sibilants /s/, /sh/, /z/, /zh/. Several approaches based on an approximation of the power spectral density (PSD) in several frequency bands were used. The experiments performed used a linear and a quadratic approximation in the frequency bands 1–3 kHz and 3–6 kHz, a cubic approximation in the frequency band 1–6 kHz and the Bézier approximation of a PSD which uses frequencies 0.3, 2, 6, 10 kHz as sampling points. The resulting parameters were used for classification of sibilants with a precision of 97.4% and with a precision of 99.4% for classification of the place of articulation.