This directory contains results for 8-component MAR with a simplified (silence removed) WSJ.
(containing 20 speakers, and 6 additional speakers for bgd model).

/data/isip002/research/nonlinear/mix_ar/mar/exp_008_order_x/ (x=0,1,2)

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_008/det_mar8.jpg

The result for MAR order 0 is reasonably good, but this is the same as a GMM.
Performance degrades for higher orders, possibly because of bad training (perhaps due to bad initialization?).