On this page, we provide several audio excerpts that illustrate the
source separation experiments presented in the paper. The songs are
extracted from the Demixing Secrets Database (DSD100).
Oracle scenario
We note that the "Isolated unwrapping" estimates are corrupted with
artifacts. We don't here much differences between the other techniques,
except for the bass part which is severely deteriorated by the Wiener
filtering and Consistent Wiener filtering estimates, while the proposed
estimator leads to a cleaner result.
Song: "One Minute Smile" by Actions. Mixture:
Bass
Drum
Other
Vocals
Original sources
Wiener filtering
Consistent Wiener filtering
Isolated unwrapping
Proposed Anisotropic Wiener filtering
Non-oracle scenario
The "Isolated unwrapping" estimates are highly corrupted with artifacts,
leading to a poor audio quality. While the Consistent Wiener filtering
technique leads to increase the quality over traditional Wiener
filtering, we do not hear any significant difference between the
consistent Wiener filtering estimates and the proposed anisotropic
Gaussian-based estimates. However, our method is 6 times less
computational costly.