Experimental Setup

We now consider an example to illustrate the algorithm described above. We again use the mixing parameters of the previous example, namely:

source	$a_i$	$\delta_i$
1	1.005	-6.5e-5
2	0.995	-2.5e-5
3	0.960	2.0e-5

For this example, we trained distributions assuming all three possible source pairings, and each time allowed 39 values of $r$. The first 19 values were uniformly spaced on the range $[0.05, 0.95]$, the next 19 were the inverse of those values, and $r=1$ was also included. It may be shown that this spacing represents the distribution that may expected on $r$ when $\vert S_u\vert$ and $\vert S_v\vert$ are uniformly distributed. For each $r$ value, 1000 values of the phase offset $\theta$ were used to create each histogram, each of which used 30 bins. This was done for each of 511 STFT frequency bins corresponding to positive (and nonzero) frequencies.⁴

For testing, we used the same 39 values of $r$ as used in the training. ⁵ We did, however, choose 11 test values of $\theta$, only one of which was coincidental with those used in the training. This was done for each positive STFT frequency bin.

We also used the DUET system on this input data. The results for each system are presented and discussed in the next subsection.