Sparse audio coding via targeted dithering and combinatorial decoding

We present a novel paradigm for sparse audio signal coding. After annihilating unperceivable components in some transform domain, the encoder buffers the resulting sparse vector into small non-overlapping frames. In each frame, the active elements' amplitudes are quantized, and with the help of a priori known unquantized “filler” vectors (whose values are placed in inactive positions), their position is encoded such that a certain function f of the filled vector is nearly integer valued. For this purpose, the quantized values of the sparse frames are pre-adjusted in a controlled manner with distortion in mind (hence the name “targeted dithering”). The decoder then progresses through the possible combinations of the nonzero elements, and verifies with the filler vector which of these combinations produces an integer valued f, thereby retrieving the active elements' positions. In preliminary tests, good quality can be obtained by encoding 44.1 kHz signals with less than 50 kbps.