Sample-by-sample multi-transform processing of nonstationary speech signals

We develop algorithms that use orthogonal transforms for sample-by-sample representation of speech signals. Two implementation structures are proposed, namely, the gradient and the Gauss-Seidel structures. The performance of these structures is found to be considerably superior to the performance of previously developed block-by-block implementation structures when low numbers of dominant components (DC) are selected. The look back recursive residual projection (LBRRP) strategy gives good results for unvoiced and transient speech segments, while the harmonics of pitch frequency (HPF) and modified spectral peaks (MSP) strategies demonstrate better performance for voiced speech segments. The major disadvantage of the LBRRP strategy is the large number of computations required.