The paper presents a new idea for coding of audio signals
whose short-time spectrum is sparse, i.e. the energy is concentrated
within few narrowband peaks. Typical examples
are solo instrument parts with distinct harmonics as well as
singing voices, especially in high registers. Such signals are
often degraded by standard perceptual codecs that assume
quite different spectral energy distribution. Our approach
uses spectral peak detection and tracking to follow the energy
bands evolving in TF plane. Individual bands are isolated
and shifted down in frequency so that they are well
approximated by only few MDCT coefficients. Moreover, we
interleave the transform coefficients in order to exploit their
similarities across bands and to increase the coding efficiency.
Experimental results show that for signals with
sparse spectra a significant improvement of coding efficiency
is achieved. This technique may be considered as a
complementary coding tool for the MPEG-4 HE-AAC codec.