Adaptive wavelet subband coding for music compression

This paper describes modelling of the coefficient domain in wavelet subbands of wideband audio signals for low-bit rate and high-quality compression. The purpose is to develop models of the perception of wideband audio signals in the wavelet domain. The coefficients in the wavelet subbands are quantized using a scheme that adapts to the subband signal by setting the quantization step size for a particular subband to a size that is inversely proportional to the subband energy, and then, within a subband, by modifying the energy determined step size as inversely proportional to the amplitude probability density of the coefficient. The amplitude probability density of the coefficients in each subband is modelled using learned vector/scalar quantization employing frequency sensitive competitive learning. The source data consists of 1-channel, 16-bit linear data sampled at 44.1 kHz from a CD containing major classical and pop music. Preliminary results show a bit-rate of 150 kbps, rather than 705.6 kbps, with no perceptual loss in quality. The wavelet transform provides better results for representing multifractal signals, such as wide band audio, than do other standard transforms, such as the Fourier transform.