Efficient vector quantisation of wideband LPC parameters using the ML-SSVQ
In this paper, we investigate the ML-switched split vector quantiser (ML-SSVQ), which uses the concept of multiple survivor paths to improve the rate-distortion (R-D) performance of conventional SSVQ at the cost of increasing the computational complexity. Despite the SSVQ being a state-of-the-art vector quantiser for coding line spectral frequencies, it suffers from high memory requirements. This can be alleviated by splitting the vectors into more parts though this comes at the cost of degrading R-D efficiency. We will show via wideband LSF experiments that the ML-SSVQ incurs less spectral distortion and outlier frames than conventional SSVQ at the same bitrate. These improvements have allowed the six-part ML-SSVQ to be competitive when compared with five-part SSVQ and the 46 bits/frame split-multistage vector quantiser with moving average predictor from the ITU-T G722.2 AMR-WB speech codec.
Proceedings of the Griffith School of Engineering Research Conference