Multiobjective Optimization Training of PLDA for Speaker Verification

1. prepare data, make directory ./data and ./temp
   put NIST SRE14 i-vector challenge offical data on "./data/", there are "development_data_labels.csv, dev_ivectors.csv, ivec14_sre_segment_key_release.tsv, ivec14_sre_trial_key_release.tsv, model_ivectors.csv, target_speaker_models.csv, test_ivectors.csv"

2. run ./python/sre14_preprocess.py.
   It will generate "./temp/sre14.mat"

3. run ./matlab/gplda_demo.m
   The script will read "./temp/sre14.mat", and the results are
   2.347, 2.456 (Development dataset, EER), 2.307 (Evaluation dataset, EER),
   0.264, 0.269 (Development dataset, MDCF), 0.261 (Evaluation dataset, MDCF).

4. run ./matlab/moplda_demo.m
   The script will read "./temp/sre14.mat", and the results are
   2.040, 2.193 (Development dataset, EER), 1.931 (Evaluation dataset, EER),
   0.233, 0.239 (Development dataset, MDCF), 0.229 (Evaluation dataset, MDCF).

5. some experiment results.

A. train lambda with development and train vectors

0. factor experiment,
   A EER, D EER, E EER, A DCF, D DCF, E DCF, factor
1. 2.531, 2.794, 2.354, 0.272, 0.277, 0.267, 1.1
2. 2.554, 2.825, 2.336, 0.269, 0.272, 0.266, 1.2
3. 2.456, 2.677, 2.176, 0.250, 0.255, 0.247, 1.3
4. 2.331, 2.579, 2.199, 0.238, 0.241, 0.235, 1.4
5. 2.207, 2.399, 2.099, 0.233, 0.235, 0.230, 1.5
6. 2.082, 2.272, 1.940, 0.230, 0.235, 0.225, 1.6
7. 2.040, 2.193, 1.931, 0.233, 0.239, 0.229, 1.7
8. 2.057, 2.180, 1.973, 0.238, 0.244, 0.232, 1.8
9. 2.136, 2.241, 2.049, 0.244, 0.250, 0.238, 1.9
10. 2.108, 2.261, 1.995, 0.242, 0.249, 0.237, 2.0

B. train lambda with development vectors

0. factor experiment,
   A EER, D EER, E EER, A DCF, D DCF, E DCF, factor
1. 2.182, 2.426, 1.994, 0.241, 0.244, 0.237, 1.1
2. 2.257, 2.438, 2.099, 0.237, 0.243, 0.232, 1.2
3. 2.369, 2.487, 2.225, 0.240, 0.245, 0.236, 1.3
4. 2.307, 2.426, 2.210, 0.238, 0.245, 0.233, 1.4
5. 2.481, 2.610, 2.330, 0.264, 0.273, 0.257, 1.5
6. 2.340, 2.425, 2.254, 0.253, 0.260, 0.246, 1.6
7. 2.347, 2.487, 2.267, 0.253, 0.260, 0.247, 1.7
8. 2.337, 2.395, 2.301, 0.264, 0.272, 0.257, 1.8
9. 2.406, 2.456, 2.351, 0.266, 0.274, 0.260, 1.9
10. 2.506, 2.549, 2.435, 0.275, 0.281, 0.269, 2.0

C. train lambda with development and train vectors

0. dimension experiment,
   A EER, D EER, E EER, A DCF, D DCF, E DCF, dimension
1. 3.522, 3.556, 3.505, 0.498, 0.507, 0.491, 50
2. 2.306, 2.329, 2.267, 0.288, 0.294, 0.284, 100
3. 2.032, 2.241, 1.919, 0.239, 0.245, 0.234, 150
4. 2.040, 2.193, 1.931, 0.233, 0.239, 0.229, 200
5. 2.070, 2.260, 1.918, 0.234, 0.240, 0.229, 250

PS:

1. The sGPLDA demo was downloaded from https://github.com/wangwei2009/MSR-Identity-Toolkit-v1.0
2. Anaconda3, Python3, require sklearn
3. Matlab R2016a
4. A EER: NIST SRE14 i-vector challenge all data EER
   D EER: NIST SRE14 i-vector challenge Development dataset EER
   E EER: NIST SRE14 i-vector challenge Evaluation dataset EER
5. A DCF: NIST SRE14 i-vector challenge all data MinDCF14
   D DCF: NIST SRE14 i-vector challenge Development dataset MinDCF14
   E DCF: NIST SRE14 i-vector challenge Evaluation dataset MinDCF14