Lossless wideband audio compression: prediction and transform

Lossless Wideband Audio Compression: Prediction and Transform This thesis studies lossless audio compression. In the domain of lossless compression, research takes place on two broad development sections, signal modeling and coding algorithm. The former is concerned with the understanding of the source signal, while coding is the more tightly specified task of efficiently representing a single symbol as a code. The focus of this thesis is the evaluation and the development of signal modeling techniques for lossless compression. Related with the modeling method used to decorrelate a signal, the data compression schemes are generally divided in two categories, predictive modeling and transform-based modeling. In the thesis, all two categories are investigated in depth and handled from the lossless viewpoint. The first contribution of the thesis is an exploration of the general audio compression systems including the lossy compression system. In predictive modeling, the structures of various linear prediction filters are introduced by presenting the fundamental autoregressive modeling. The prediction filters including the approaches to the nonstationary signal modeling and to the adaptive linear prediction filters are explored and evaluated by testing within a prototypical lossless audio compression system. For transform modeling, two well-known subband transform coding methods, Laplacian pyramid and subband coding scheme, are first described, and then the design methods of perfect reconstruction multirate filter banks are studied. Concerning with the modulated lapped orthogonal transform, the efficiency of linear prediction from subband and from fullband is formally compared and empirically examined. Wavelet transform is in depth studied from the various viewpoints in order to find the theoretical relationship between the wavelet and the multirate filter banks. Theoretical and practical aspects of reversible transforms are discussed by introducing the S-transform, S+P transform, and RTS transform. The lifting method is examined as a means to realize the biorthogonal wavelets. Integer lifting scheme with rounding-off method is investigated to construct reversible version of wavelet transforms and its performance is validated by applying to lossless audio compression. Finally, some of the more important results presented in this thesis are summarized with the suggesting directions for future research.