Sparse Independent Vector Analysis Based on Mel Filter

To make independent vector analysis (IVA) robust for whitening, sparse IVA clips spectrum in high frequency bands, because whitening generates artificial noise in high frequency bands. In this paper, to avoid clipping of source spectrum by sparse IVA, we propose an application of Mel filter to the observed spectrum before clipping in order to emphasize spectrum in low frequency bands, because source spectrum in low frequency bands is often more important and sparse than that in high frequency bands. The effectiveness of the proposed method is confirmed by sound source separation experiments.