The speaker detection system we have developed for NIST SRE 2010 is a GMM Supervector - SVM system. We use Mel Frequency Cepstral Coefficients with 19 static, 19 delta and 1 delta-energy component 
( total dimension 39 ) as features. Silence removal is achieved by training a bi-Gaussian model from frame energies for each segment. Mean + standard deviation of the Gaussian with lower energy is 
used as threshold. Feature warping with 3 second window is applied after silence removal. 2048 mixture gender-dependent Gaussian Mixture Models (GMM) are trained from SRE06, SRE08 and SRE08 follow-up
databases. SRE08 databases are splitted in two parts. The first part is used in system development stages such as world model training, znorm utterances and impostor utterances. The second part is 
used for testing. Relevance MAP adaptation with a relevance factor of 8 is used for obtaining the GMM for each utterrance. The means of the world model are subtracted from the corresponding means of 
the GMM and the resulting vectors are normalized by variance and concatenated. Finally the supervectors are scaled to be unit norm. SVMTorch is used to obtain SVM models. Znorm score normalization is 
applied using utterances from SRE06, SRE08 and SRE08 follow-up databases.