Iterative monaural audio source separation for subspace grouping

Monaural blind audio source separation usually separates a mixture into more signals than active sources. Therefore, a clustering of the separated signals is needed to reconstruct the sources. We propose a new iterative clustering and show that this approach outperforms classical clustering approaches which use features of the separated signals for clustering. The iterative clustering starts with the separation into two source estimates. Based on this, at each iteration the squared error between the source estimates of the former iteration and a linear superposition of the separated signals of the current iteration is minimized. The corresponding linear superposition generates new source estimates. The algorithm is evaluated on a large test set regarding melodies of different instruments, singing, and speech from the EBU.