Circular linear prediction modeling for speech coding applications

Previously, we introduced a window-less linear-prediction method known as circular linear prediction (CLP) modeling for pitch-synchronous analysis of the speech spectrum. This paper presents the spectral estimation properties of the CLP method using various synthetic speech signals including fully-voiced, unvoiced and partially-voiced speech signals. As the true spectrum is known, it is possible to make objective measurements on how well the CLP method estimates the true spectrum. In addition, a multi-cycle generalization of the CLP method is proposed. This new method improves the spectral estimation performance of the CLP method for short pitch cycles (less than 7.5 ms). These experiments prove that the CLP method and its multi-cycle generalization are viable alternatives to the autocorrelation method for modeling speech signal, especially in transition regions.