This directory contains the convergence results for MixAR EM estimation for two speakers (011, 013) from WSJ corpus.
Code for this is in $ISIP_DATA/research/nonlinear/mix_ar/exp_001/testClassifier.cc1.

1) One large window (11 s) of speech, and ran several EM iterations over this same window. The likelihood converged with iterations, and the nan/zero problems that were troubling till last week
vanished. The result is plotted in:

http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/exp_001/wind_11s.jpg

2) With reduced window size 5s, repeated the same experiment for two
speakers to test for convergence and to see the difference in likelihoods for
classification. The plot is in:

http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/exp_001/iter_50_wind_5s.jpg

3) Made the next set of experiments more general:

 a)Recursive EM using Accumulate mode so that at each EM iteration, the
speech data comes from a different frame of data.
 b) Open loop testing: Used a test speech data (50 s) distinct from the training
data.
 c) Test convergence/performance with 3 different window size 5s, 1s, 0.5s.

The plots are shown in:
http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/exp_001/wind_5s.jpg
http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/exp_001/wind_1s.jpg
http://www.ece.msstate.edu/research/isip/projects/nsf_nonlinear/doc/baseline_results/speaker_recognition/mix_ar/exp_001/wind_halfs.jpg

While initially they seem to converge, they start behaving more erratic afterwards.