Fast implementation of an improved parametric audio coder based on a mixed dictionary

This paper deals with the application of adaptive signal models for representing transients and sinusoids at the same stage in a parametric audio coder. To accomplish such a goal, we search for sparse approximations by means of matching pursuit with a mixed dictionary, instead of using two different dictionaries that operate in cascade. In such sense, complex exponentials and wavelet packets are chosen for modeling the tonal and transient features of an audio signal, respectively. At each iteration of the pursuit, the mixed dictionary function that extracts the most energy from the residue is selected. This function will be either a complex exponential or a wavelet packet, depending on the characteristics of the residue at that iteration. Experimental results clearly show the objective (compression rate) and subjective (% preference) advantages of the mixed dictionary over two cascaded dictionaries. The approach proposed in this paper is successfully applied for parametric audio coding purposes, assuring better perceptual audio quality than MPEG2/4-AAC at 16 Kbits/s for most of the CD-quality one channel audio signals considered for testing.