Predictive VQ for bandwidth scalable LSP quantization [speech coding applications]

Predictive vector quantization (PVQ) for a bandwidth scalable LSP (line spectrum pairs) quantizer was studied. In the quantizer, LSP parameters obtained from a wide-band (16-kHz sampling) speech signal were quantized using LSP parameters encoded in a narrow-band (8-kHz sampling) layer. We designed several configurations of 16-bit quantizer utilizing PVQ and compared their performance. Simulation results showed that effective exploitation of interframe and intra-frame prediction improved objective performance. Compared with a baseline algorithm, PVQ using normal inter-frame prediction achieved better SD (spectral distortion) performance by 0.15 dB, and further improvement of 0.05 dB was gained by introducing novel interframe and intra-frame prediction. Robustness against frame erasure was considered by adopting switched prediction and "forgetting" capability. The designed memory-based PVQ algorithms were found to be robust up to 20% frame erasure rate.