This baseline build on well-established iVector/PLDA speaker verification framework for SRE 2010 female tasks. It is based on Kaldi SRE2008 example,forked under version r3473.
#There are two setups for SRE2010 with different using of SRE2008 data ##Setup 1 SRE2008 is not used for Tmatrix,PLDA,LDA training, reported as the clean iVector/PLDA base line in [1] with average EER 3.50%. Results reported by other researchers on SRE2010 in papers[?] ###UBM training Switchboard II Phase 2 and 3, Switchboard Cellular Part 1 and 2, and NIST SRE 2004, 2005 and 2006 ###I Vector Extractor training Switchboard II Phase 2 and 3, Switchboard Cellular Part 1 and 2, and NIST SRE 2004, 2005 and 2006 ###Development Dataset SRE2008 ###Test Dataset SRE2010
##Setup 2 SRE2008 is used for Tmatrix,PLDA,LDA training (WIP),results reported by other researchers in this setup[??] ###UBM training Switchboard II Phase 2 and 3, Switchboard Cellular Part 1 and 2, and NIST SRE 2004, 2005 and 2006,SRE2008 ###I Vector Extractor training Switchboard II Phase 2 and 3, Switchboard Cellular Part 1 and 2, and NIST SRE 2004, 2005 and 2006,SRE2008 ###Development Dataset XX ###Test Dataset SRE2010
#Prerequisites ##data having the Data listed above ##NIST DATA ##NIST Keys #Kaldi #Install Kaldi #Kaldi structure
If you use this baseline kindly cite paper:
[1] Steven Du, Xiong Xiao, Eng Siong Chng, "DNN FEATURE COMPENSATION FOR NOISE ROBUST SPEAKER VERIFICATION", in proceedings of ChinaSIP 2015