Skip to Main Content
We investigate in this paper the lossless audio coding performance achievable when the linear prediction is performed in the decompanded domain and the predictors are designed by the interleaved quantization-optimization Burg algorithm. The Burg algorithm decouples the optimization process into stages for optimizing separately each reflection coefficient, propagating at the same time the prediction residuals for all predictors of orders from 1 to an upper bound M. Model order selection can be thus performed utilizing precise values for both reflection coefficient codelength and for residual codelength for each predictor order. Additionally, the reflection coefficients can be further optimized with respect to the meaningful criterion, which is the codelength of the residuals, instead of the mean square residuals used in the traditional Levinson-Durbin optimization. Although achieving high order predictors becomes computationally well feasible, the experimental results show that prediction orders higher than 20 are rarely needed since the compression ratio saturates very quickly after this limit. The overall coding scheme compares favorably with the recently adopted G711-LLC standard.