Switched split vector quantiser and its application to LPC parameter quantisation in wideband speech coding

Abstract

In this paper, we present an overview of the switched split vector quantiser (SSVQ), its advantages over the traditional split vector quantiser, and its application to LPC parameter quantisation in wideband speech coding. By utilising an unconstrained switch vector quantiser as an initial stage, the SSVQ is able to exploit global vector dependencies in order to compensate for rate-distortion (R-D) losses in the split vector quantiser, which are due to the structural constraints on the codebook. The resulting scheme is not only more efficient in the R-D sense, but also possesses lower computational complexity than the split vector quantiser. We apply the SSVQ to quantising line spectral frequency (LSF) parameters of wideband speech and compare its spectral distortion performance with the split vector quantiser (SVQ), PDF-optimised scalar quantiser, and the split-multistage vector quantiser (S-MSVQ) with MA predictor from the ITU-T G.222.2 AMR-WB speech coder. The results show that the SSVQ requires less bits for achieving transparent coding of wideband LSFs than SVQ and scalar quantisers. Finally, SSVQ (which is a memoryless scheme) achieves comparable spectral distortion with the S-MSVQ with MA predictor at 36 and 46 bits/frame.