Multiresolution Sinusoidal Speech Model Using Elliptic Band Pass Filter

The sinusoidal speech model represents a speech signal as a linear combination of sinusoids with time-varying parameters {amplitudes, frequencies, and phases}. However, one drawback of this model is that the analysis window width is generally fixed in analyzing the signal. Since each sinusoidal parameter has different frequencies, an analysis window with fixed width can’t guarantee an optimum spectral resolution to each sinusoidal parameter. In this paper, we propose and implement a multiresolution sinusoidal speech model using elliptic band pass filter to overcome this drawback and to estimate the sinusoidal parameters more precisely. Experimental results have shown that the proposed model can achieve better performance than that of the classical sinusoidal model.