13
Estimation of the Fundamental Frequency of the Speech Signal Using PCC Interpolation With the BL Kernel
Authors: Zoran N. Milivojević, Zoran S. Veličković
Number of views: 38
This paper describes the construction of a parametric cubic interpolation kernel based on the blending method. The blending kernel is constructed by mixing two third-order interpolation kernels, the one-parameter Keys kernel (parameter α) and oMoms3 kernel. The proportionality of the participation of the oMoms3 kernel in blending kernel is represented by the blending factor (w). Created blending kernel has two parameters (α, w) which adjusting affects the accuracy of interpolation, that is, the reduction of interpolation error. The characteristics of the blending kernel are shown graphically in both, the time and spectral domains. After that, the algorithm for estimating the optimal parameters of the blending kernel is presented. The algorithm is described using a pseudo code. Subsequently, the precision of estimation of the fundamental frequency of the speech signal in the spectral domain, was tested experimentally, using an estimation algorithm. First, the speech test signal is processed in the time domain using some window (Hamming, Hann, Kaiser and Triangular). Subsequently, the speech test signal was transformed into the spectral domain using fast Fourier transformation (FFT). Then, using Peak-picking algorithm, a dominant spectral component (fundamental frequency) was determined. Most often the real fundamental frequency is between the two dominant spectral components. The real fundamental frequency is determined by applying a parametric convolution with a blending interpolation kernel. The estimation precision is represented by the mean square error (MSE) between the estimated and the real fundamental frequency. The optimal kernel parameters are determined by minimizing the mean square error, and the appropriate window is selected. The results are presented by tables and graphics.