We now consider the performance of the Bayesian Two Source Modeling (BTSM) approach in the above example. To make clear the strengths of the new approach, we first present a table summarizing and comparing the DUET and BTSM approaches. A second table gives numbers showing performance relative to the claims made in the first table.
| Issue | DUET | BTSM | |
| For up to one source active: | Identify and demix active source | Yes | Yes |
| For up to two sources active: | Identify at least one active source | Implicit, medium accuracy | Explicit, high accuracy |
| Identify and demix both active sources | No | Yes | |
| Identify relative weighting of active sources | Implicit, low accuracy | Explicit, high accuracy | |
| Possible to use source knowledge to inform source ID? | No | Yes | |
| Issue | DUET | BTSM | |
| For up to one source active: | Identify and demix active source | 100% | 100% |
| For up to two sources active: | Identify at least one active source | 86% | 100% (trivial 3 source case) |
| Identify and demix both active sources | 0% | 59% | |
| Identify relative weighting of active and non-active sources | 20% average error | 13% average error | |
| Possible to use source knowledge to inform source ID? | No | (Yes, though not done here) | |
The tables help highlight some important issues. First, the BTSM system is only beneficial when two sources are active. If in fact the sources observed obey W-disjoint orthogonality (have no overlap in time-frequency space), the BTSM system does not help, other than via the Bayesian framework for introducing prior knowledge about the sources. The ability of the BTSM system to gauge relative weighting of at least one source even when the initial guess is wrong is only beneficial in the two source case. Hence, the more often two sources are present, the more beneficial the system is.
When compiling the statistics above, we created a situation where exactly two of the three sources were present, and each was modeled as independent and identically distributed with a uniform distribution on amplitude and phase. This may or may not represent reality, though it is not possible to represent every realistic case.
The mixing parameters and source
distributions will both bear on the results achieved by the
system. Specifically, the more often extreme values of 
are seen, the more
often the DUET system will produce low error estimates of relative weights. When
the 
, representing equal source amplitudes, the DUET system error will
always be at best 50% for the active sources. Thus, the current test
scenario is intended only as an illustration of potential.
Future work will include specific musical examples. In cases of independent and sparse speech sources such as those for which the DUET system was originally intended, two source overlap is rare. There, the BTSM improvement is of modest significance, chipping away at an error bound of about 15 percent in the four speaker case, for example.
The accuracy rates of 100% in the table above represent only the performance of the systems in the current hypothetical noiseless, anechoic environment. Future testing will seek to determine how the current approach performs in more realistic situations.
Errors are still possible with the BTSM approach. If the wrong source combination is guessed, however, we see that often the relative amplitude of the louder source is correctly or near-correctly guessed. This issue manifests itself differently in the DUET system. There, one source is always chosen, and given a default weighting of 100%. If two sources are present, the weighting of the guessed source cannot be 100% (and may be zero) and an error equal to the difference is created. We see that this error is often larger than that generated by the BTSM system even when its guesses are otherwise incorrect.