Vector quantizationwith renormalized splits for wideband speech

Abstract

A method is proposed for split vector quantization (SVQ) that, while keeping up with the lower search complexity inherent to SVQ, reduces the attendant split loss. It operates partially in the training and encoding phases by normalizing the bandwidth covered by each split. The action is completed in the decoding phase by renormalizing the bandwidth spanned by each split of the codebook to that allowed by a nonoverlapping constraint with neighboring split bands. Therefore, the method is referred to as renormalized SPV (RSVQ). The performance of RSVQ is investigated in comparison to standard SVQ for coding line spectral frequency (LSF) vectors that parameterize the spectral envelope of wideband speech and it is found to save four bits, reaching transparent coding at 42 bit/frame.