Vector quantization and perceptual criteria in SVD based CELP coders

Experiments in which singular value decomposition (SVD) is applied to code excited linear prediction (CELP) coders are presented. This means that the concept of selection of the optimum innovation pattern through an analysis by synthesis (ABS) method can be replaced by a weighted mean-squared-error computation in a transformed domain. The transformed signal is obtained by representing the residual signal on the orthonormal basis obtained by SVD of the matrix of the truncated linear predictive coding (LPC) filter´s impulse response. The weighting consists of the vector gain and the singular values (SVs) of the matrix. The characteristic of the SVD approach is that the excitation signal is represented as a linear combination of orthonormal signals whose spectra show characteristics quite similar to those of bandpass filters. Moreover, errors in the amplitude of each component at the filter input reflect errors only in the corresponding components at the filter output, weighted by the associated SV. This characteristic can be exploited by incorporating a simplified auditory model to determine the subjective importance of the singular value components