AUTOREGRESSIVE AND CEPSTRAL PARAMETRIZATION IN HARMONIC SPEECH MODELLING

Anna Madlová

   Two methods for coding the amplitude and minimum-phase spectra of the speech signal are compared. Autoregressive (AR) and cepstral parametrizations of the same spectral envelope use the same number of parameters. Computational complexity of the analysis is compared for the two methods. Using both AR and cepstral parametrization the harmonic synthesis is compared for concatenation of pitch-synchronous frames and overlap-and-add (OLA) of consecutive pairs of the same pitch-synchronous frames. Quality of synthesis is compared using the RMS log spectral measure. The mean value of the spectral measure is rather similar for all the four combinations of analysis and synthesis methods. However, the standard deviation of the spectral measure is lower for the OLA synthesis.

Keywords: harmonic speech model, autoregressive and cepstral parametrization

[full-paper]