Advantages Over the DUET System

The model in the DASSS system is advantageous over the DUET system in several circumstances. First, we recall that in cases where more than one source is active, the DUET system will produce a source estimate in accordance with equation 1 without any knowledge of the fact that more than one source is active. This can produce estimates of intermediate sources in the parameter space without warning. Using the DASSS function of equation 6, however, it is possible to see that two sources are active when no single-source model appears to be appropriate. This indication that two sources are active provides an additional benefit. Given knowledge that two more sources are present, the system can proceed to attempt to fit multi-source models to the data. This is especially important when the data does not obey W-disjoint orthogonality. Music ensemble signals do not obey this assumption, because the harmonic structure of music and the harmonic nature of instruments and voices leads to frequency bins being shared by multiple sources. Approaches to this case are suggested in section 5. Another benefit from DASSS occurs when one source is strongly active in a particular frame's bin, but a much weaker source is also active. Again, the DUET system will produce an estimate of the active source from 1 which could produce spurious indications that a third source is active, or that no particular source is active. The DASSS system, however, will show that the error for the hypothesis of source $g$ is much less than that for any other source model. In such cases, the system may choose to assign the energy to the stronger source, rather than to an incorrect third source or no source at all.