Aspects of melisma synthesis

Abstract

The problem of restoration of the comparatively long fragments of songs called melisma was proposed to be cast in a framework of a signal synthesis. The paper investigates several possibilities to synthesize melisma employing original, approximated and modeled parameters of melisma. Several synthesis ways are explored: employing linear prediction coding (LPC) coefficients and original glottal signal, and LPC coefficients, modeled glottal signal, gain and pitch. New method for synthesis of melisma employing fragment of original glottal signal and original LPC coefficients, gain and pitch is proposed. In order to cope with the variability of speech signals simple but effective procedure based on the cross-correlation test is proposed.