A novel application of ARMA modelling to audio coding

In many transform audio codecs, the data set to encode consists of groups of transformed coefficients located in frequency bins. In this paper, we propose to use certain Autoregressive Moving-Average (ARMA) modelling techniques to encode the transformed coefficients by taking into account their perceptual significance. The novelty lies in the fact that the modelling is performed with the transformed coefficients viewed as the constrained variables of an optimization problem. This differs fundamentally from the usual methods consisting of weighing certain terms in a cost function. We describe situations where satisfactory fits can be obtained at low model orders if some errors are tolerated. Accessorily, we introduce a permutation-based scheme which allows increasing the output precision with very small information overhead. The ideas presented are then tested on the MDCT coefficients of real data as part of a simple audio codec, confirming the potential of the method.