Efficient coding of linear predictive coefficients for wideband speech

This paper considers the coding of the linear predictive (LP) filter coefficients for applications where the highest quality natural-sounding speech reproduction is required. Bandwidths as high as 16 kHz and linear prediction orders of 38 are considered. To deal with the increase in model order, a method is proposed to code the trajectory of the line-spectral pair (LSP) representation of the LP coefficients. To improve the performance of the system, a hybrid scheme is proposed which utilises a combination of intra-frame LSP coding and trajectory-based coding. This scheme exploits the reduced sensitivity of the human auditory system to distortion at higher frequencies.