Multi Channel Audio Processing: Enhancement, Compression and Evaluation of Quality (Aachener Beiträge zu digitalen Nachrichtensystemen)

The demand for high definition audio and video services is rapidly increasing. Two representative examples for this are audio-visual conferencing or video storage and delivery. In this context, efficient techniques are required for enhancement and compression of multi channel audio signals with compatibility to, e.g., mono or stereo systems. In this thesis, novel signal processing algorithms for both enhancement and compression of multi channel signals are developed and theoretical performance bounds are derived. Additionally, a novel instrumental quality measure for the evaluation of multi channel signal processing algorithms is proposed.
Enhancement schemes for both the recording and the reproduction side are introduced. This includes the optimization of a near field filter-and-sum beamformer to achieve a target directivity characteristic at the recording side. For the reproduction side, an efficient postfilter is presented which increases the speech intelligibility by taking the positive influence of early room reflections into account.
The main part of this thesis covers multi channel predictive compression of audio signals. A predictive multi channel coding system is presented and analyzed. Performance bounds are derived and two methods for an adaptive bit rate distribution between inter channel and intra channel prediction are devised. Novel multi channel noise shaping concepts are introduced. The performance of the compression system is quantified by instrumental measures.
A novel instrumental measure is introduced for the evaluation of multi channel signal enhancement and compression. It combines the proven single channel quality measure PEAQ with a binaural auditory model and a mathematical model of cognitive behavior, providing a reliable evaluation of quality perception and spatial fidelity. The inclusion of spatial information into the instrumental quality measurement leads to a consistently high correlation between the instrumental measure and a listening test.