This directory contains results for 8-component MAR with our modified NIST database in:
$ISIP_DATA/research/nonlinear/mix_ar/mar/exp_002/

The speaker recognition result is in:

http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/mar/exp_002/det_mod_nist_08.jpg

Here again, at the EER point, similar to the sustained phone database, the false alarm prob is relatively much lesser compared to the miss probability, which is quite high.

As for EM convergence:

http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/mar/exp_002/em_4516.jpg
for speaker 4516 has not converged well within 30 iterations,

while for speaker 5063,
http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/mar/exp_002/em_5063.jpg
reasonable convergence is reached within 30 iterations.