Spectral envelope quantization based on conditional inter-frame prediction

Abstract

The human vocal tract system is commonly modeled by a linear predictive coding (LPC) filter whose coefficients are transformed into a line spectral frequency (LSF) vector for quantization. Predictive split-vector quantization (PSVQ) based on an auto-regressive model (AR-PSVQ), which exploits the inter-frame correlation of the LSF vectors, provides a better rate-distortion performance compared with quantization methods that only consider the intra-frame correlation. In the proposed conditional PSVQ (C-PSVQ), the conditional distribution of the current-frame LSF given the previous-frame LSF is taken into account. Compared with AR-PSVQ, C-PSVQ gains 1 bit in terms of average spectral distortion and 2 bits in terms of the number of outlier frames. Memory requirements and computational complexity of C-PSVQ are similar to those of AR-PSVQ.