Multiobjective Optimization Training of PLDA for Speaker Verification \

1. prepare data, make directory ./data and ./temp
   put MCE18 offical uncompressed data on "./data/",
   there are "bl_matching.csv, trn_blacklist.csv, trn_background.csv, dev_blacklist.csv, dev_background.csv, dev_bl_id.ndx, tst_evaluation.csv, tst_evaluation_keys.csv"
   open the website (http://www.mce2018.org/) for data requirement.

2. system: length-normalization + LDA + PLDA (MotPLDA) + score-normalization.

A. run ./python/mce18_plda_preprocess.py
It will generate "./temp/mce18.mat"

Option 1:
B1. run ./matlab/gplda_demo.m for PLDA
The script will read "./temp/mce18.mat", and it will generate "./temp/mce18_result.mat"

C1. run ./python/mce18_plda_eval.py
The script will read "./temp/mce18.mat", and the results are
Test set score using training and development set :
Top S detector EER is 6.75%
Top 1 detector EER is 9.39% (Total confusion error is 270)

Option 2:
B2. run ./matlab/moplda_demo.m for MoPLDA
The script will read "./temp/mce18.mat", and it will also generate "./temp/mce18_result.mat"

C2. run ./python/mce18_plda_eval.py
The script will read "./temp/mce18.mat", and the results are
Test set score using training and development set :
Top S detector EER is 5.41%
Top 1 detector EER is 7.32% (Total confusion error is 204)

Reference:
[1] Suwon Shon, Najim Dehak, Douglas Reynolds, and James Glass, “Mce 2018: The 1st multi-target speaker detection and identification challenge evaluation (mce) plan, dataset and baseline system,” in ArXiv e-prints arXiv:1807.06663, 2018.

[2] L. He, X. Chen, C. Xu, and J. Liu, “Multiobjective Optimization Training of PLDA for Speaker Verification,” ArXiv e-prints arXiv:1808.08344, Aug. 2018.

[3] L. He, X. Chen, C. Xu, and J. Liu, “Multiobjective Optimization Training of PLDA for Speaker Verification,” submitted to ICASSP 2019.

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

He Liang, [email protected]
Oct. 30, 2018