Distortion components in audio source separation

In the source separation process, the estimate ŝj of a target source sj is obtained from the mixture x=s1+...+sJ of several sources.
The distortion is modeled as the sum of three distortion components:
ŝj-sj=eTarget+eInterf+eArtif
We propose a new decomposition method to estimate eTarget, eInterf and eArtif. Here are some sound examples to illustrate the resulting decomposition and compare it to the state-of-the-art decomposition method (E. Vincent, H. Sawada, P. Bofill, S. Makino, and J.P. Rosca, First Stereo Audio Source Separation Evaluation Campaign: Data, Algorithms and Results, Int. Conf. on Independent Component Analysis and Signal Separation, 2007).

Example 1

Clean target (sj):
Mix (x):
Estimate (ŝj):
Components Proposed decomposition State-of-the-art decomposition (SASSEC/BSS eval)
Distorted target (sj+eTarget)
Target distortion (ejTarget)
Interference (eInterf)
Artifacts (eArtif)
Go to: Example 2 Example 3 Example 4 Example 5 Example 6

Scatter plots

The proposed decomposition vs state-of-the-art decomposition are compared on a large number of test sounds, using several criteria:
Scatter energy ratios Scatter energy ratios
Scatter audio q

Clean targets, mixtures and estimates are under this license statement. Distortion component sounds are made available under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 2.0 license.

Go to: Example 2 Example 3 Example 4 Example 5 Example 6