Discrete cosine transform based de-noising of glottal pulses

Abstract

Reliable estimates of the glottal function are of major importance in speech/voice processing for the characterization of voicing conditions, description of various phonation types, and identification of their parameters. This paper presents a new method for de-noising glottal wavelets (Differentiated Glottal Volume Velocity Pulses) and separation of the noise component, based on an approximation of their Discrete Cosine Transform as a sum of Exponentially Damped Sinusoids. The identification of the Exponentials' parameters leads to convenient estimation of ''clean'' glottal wavelets and thus separation of noise disturbances. The method is compared to standard Low-pass filtering and Wavelet de-noising using Monte Carlo simulations on synthetic Liljencrants-Fant glottal pulse models. As shown, the method supercedes for lower SNRs. Moreover, the method does not require exact determination of control parameters thus offering ease of implementation.