New filtering scheme for split-band voice coding

Abstract

This paper presents a new filtering scheme for spliting speech signals into sub-bands which results in no aliasing problem when the sub-band signals are sub-sampled and recombined. The high-pass and low-pass filters are an odd-length complementary linear phase filter pair. As a consequence, the high-pass output can be obtained by subtracting the low-pass filter output from the appropriately delayed input. This produces a factor-of-two reduction in the computation load when compared to the conventional even-length quadrature mirror filter. Another factor-of-two saving in computation load can be gained by using half-band low-pass filters whose every other coefficient is zero. Furthermore, the filters can be designed using well-known minimax optimisation techniques such as the Remez exchange algorithm and integer programming.