Low-complexity wideband LSF quantization by predictive KLT coding and generalized Gaussian modeling

In this paper we present a new model-based coding method to represent the linear-predictive coding (LPC) parameters of wideband speech signals (sampled at 16 kHz). The LPC coefficients are transformed into line spectrum frequencies (LSF) and quantized by switched AR(1)/MA(1) predictive Karhunen-Loeve transform (KLT) coding. Compared to previous work, the main novelty lies in the use of improved quantization to represent the (transformed) prediction error of LSF parameters. Generalized Gaussian modeling is applied for this purpose. We review existing methods to fit the free model parameter of a generalized Gaussian model to real data and show that that the distribution of the prediction error for LSF parameters is indeed very close to Laplacian. Experimental results show that the proposed LSF quantization method has a performance close to classical vector quantization (AMR-WB LPC quantization) at 36 and 46 bits per frame with a much lower complexity for both design and operation.