Efficient spectral magnitude quantisation for high-quality sinusoidal speech coders

An efficient method for quantising the magnitude of the LP-residual spectrum is proposed for high quality sinusoidal speech coding. It exploits the intra-frame perceptual preference of the magnitudes using mel-scale-based warping, and the inter-frame statistical correlation of the successive magnitude vectors using predictive coding. A switched quantiser between memoryless and predictive coding schemes is also introduced for the resilience against channel errors while producing performances, in terms of weighted signal to noise ratio, superior to a memoryless quantiser and comparable to a predictive quantiser. A subjective speech quality test shows its suitability in developing high-quality sinusoidal speech coders.