MATLAB implementation of DBN-based Speaker Verification using iVector and GMMUBM on the NIST04 and NIST08 dataset in this code, we use Deep Belief Network as post-processing in the speech feature extraction stage to generate more efficient features using MFCC as DBN input.
**** Trained GMM-UBM and iVector models will be shared after acceptance of the paper ****
we used MATLAB toolboxes:
- DeeBNet
- MSRIdentityToolkit
dbn.m is our proposed deep belief network that performs the post-processing in the DBNFea.m that is called in postProcessingFeatures.m we train this net using the DeeBNet toolbox.
gmmUbmSpeakerVerification.m is prepared to use GMMUBM for speaker verification. iVectorSpeakerVerification_jfa.m is prepared to use iVector and JFA for speaker verification.
It is necessary to mention that we used HTK for MFCC feature extraction and SSVAD [1] to remove silence in Linux.
[1] M.-W. Mak, and H.-B. Yu, “A study of voice activity detection techniques for NIST speaker recognition evaluations,” Computer Speech & Language, vol. 28, no. 1, pp. 295-313, 2014.