Towards weighted mean-squared error optimality of scalable audio coding

This dissertation is concerned with efficient bit rate scalability of compression algorithms that optimize a weighted squared error (WSE) distortion metric. The major objective is to minimize the unavoidable performance loss incurred by encoding schemes in allowing for successive refinement of the bit stream. The conventional multi-layer approach to scalability incurs high performance penalty when encoding modules operate at low rate and the distortion metric differs from the simple mean-squared error. This fact motivates us to design efficient fine-grain scalable compression schemes suitable for a perceptually motivated objective metric such as the noise-mask ratio employed in audio compression systems. The ultimate goal is to demonstrate the practical application of proposed schemes via scalable compression of the audio signal.; The dissertation first focuses on system analysis of conventional multi-layer approaches to scalable compression so as to identify the cause of suboptimality. It then attacks the problem of scalability independently on three distinct and mutually exclusive fronts: quantization, prediction, and selection of encoding parameters. The subsequent combination of the proposed approaches is shown to yield a superior scalable coder at minimal increase in the overall computational cost.; The first part of the dissertation is concerned with scalability for the memoryless entropy coded scalar quantizer (ECSQ) under the optimization of the WSE metric. By considering the compandor representation of the ECSQ, it is demonstrated that asymptotic (high resolution) optimal scalability in the operational RD sense is achievable by quantizing the reconstruction error in the compandor's compressed domain. This work is then fundamentally extended to the case of low rate quantization by the use of a conditional enhancement-layer quantization (CELQ) scheme where the enhancement-layer quantizer is switched depending on the base-layer parameters. Given the above memoryless quantization scheme, the second part of this dissertation attacks scalable compression of sources with memory. The switched estimation-theoretic (sET) prediction scheme is proposed wherein an enhancement-layer signal estimate is derived in a manner that exploits all the information available at that layer. (Abstract shortened by UMI.)...