Inverse filtering for virtual acoustic imaging systems
The research topic of this thesis is the use of inverse filtering for the design and implementation of two-channel virtual acoustic imaging systems that utilise loudspeakers. The basic objective of such systems is to invert the electroacoustic plant between the input to the loudspeakers and the output at the listener’s ears and hence make it possible for a pair of binaural signals to be locally reproduced at the position of the listener’s ears. As a starting point for the research presented, a previously introduced type of inverse filtering design is considered in which the inverse is implemented with FIR filters. The basic formulation of this design is described and a number of innovative points regarding its implementation are made. An experimental procedure is then formulated for the evaluation of the effectiveness of this inverse filtering design that is based on objective measurements of the inversion process. Unlike previously employed methods that are based on computer simulations or subjective experiments, the introduced experimental procedure is shown to be very efficient in isolating and exactly quantifying the effect on the accuracy of the inversion of a number of errors and approximations typically present in the implementation. A detailed evaluation is thus presented of the inverse filtering design at hand in realistic conditions of implementation. Subsequently, a novel method for the off-line implementation of the inverse filtering is presented that utilises recursive filters of lower order. In this method, the responses of the inverse filters are decomposed into two parts, one realisable in forward time and one in backward time. The effectiveness of this new method for the implementation of the inverse is tested and compared with a small selection of the objective evaluation results described above. Finally, an algorithm for the on-line implementation of the forward-backward inverse filtering is proposed and its computational cost is compared with the currently available frequency-domain block-processing filtering algorithms.