Comparison of Signal Reconstruction Methods for the Azimuth Discrimination and Resynthesis Algorithm

摘要

The Azimuth Discrimination and Resynthesis algorithm, (ADRess), has been shown to produce high quality sound source separation results for intensity panned stereo recordings. There are however, artifacts such as phasiness which become apparent in the separated signals under certain conditions. This is largely due to the fact that only the magnitude spectra for the separated sources are estimated. Each source is then resynthesised using the phase information obtained from the original mixture. This paper describes the nature and origin of the associated artifacts and proposes alternative techniques for resynthesising the separated signals. A comparison of each technique is then presented.