Asymptotically Optimal Model Estimation for Quantization
We present a low-delay, constrained-entropy, backward adaptive, linear-predictive audio coder with low computational complexity. In contrast to most practical linear-predictive coders, the coder facilitates the exploitation of reverse waterfilling. The coder uses time-invariant quantization step size and constrained-entropy coding, thus eliminating the convergence problems of backward adaptation near signal transitions. Yet rate variations are kept small by the usage of a mixture model density for the signal. The mixture model has the backward adapted model and a second model as components and the component probability is transmitted. Experimental results confirm the advantages of the coder structure and show that the coder provides good overall performance.