Automatic detection of optimal azimuth widths for sound source separation using Adress

The Adress algorithm has proved successful as a means of separating sound sources from stereo mixtures. The algorithm has two main parameters, azimuth position and azimuth width, and these are typically set by the user for each source individually. For the separation of large amounts of audio material, such as in an audio archive, a method of automatically setting these parameters would be of use. This paper proposes a method of automatically obtaining the azimuth widths for the sources in a mixture by balancing the reconstruction error between the original spectrograms and the resynthesised spectrograms with the sparseness of the recovered sources, using an L-curve type approach.