|121st AES Convention
|San Francisco, USA
|Juan José Burred, Thomas Sikora
|Comparison of frequency-warped representations for source separation of stereo mixtures
|We evaluate the use of different frequency-warped, nonuniform time-frequency representations for the purpose of blind sound source separation from stereo mixtures. Such transformations enhance resolution in spectral areas relevant for the discrimination of the different sources, improving sparsity and mixture disjointness. In this paper, we study the effect of using such representations on the localization and detection of the sources, as well as on the quality of the separated signals. Specifically, we evaluate a constant-Q and several auditory warpings in combination with a shortest path separation algorithm and show that they improve detection and separation quality in comparison to using the Short Time Fourier Transform.