System Description

Our submission for the 2010 NIST SRE used two SVM-based acoustic systems, GLDS-SVM and GSV-SVM, combined by a weighted average of their respective scores.

Front-end
For both systems, we used 15 Perceptual Linear Prediction (PLP) coefficients along with log-energy plus all first and second order derivatives, making up a total of 48 features per vector. Vectors were computed from the speech signal, previously pre-emphasized and bandpass-filtered from 30Hz to 3400Hz every 10ms and within a 30ms sliding window. Except for the energy coefficients, feature warping was applied on all other features using a single Gaussian and a window length of 3s. Speech/non-speech segmentation was performed by thresholding the energy coefficient. The threshold was set so that the 30% of the frames were classified as speech.

GLDS-SVM System
The GLDS-SVM system uses a simplified Generalized Linear Discriminant Sequence (GLDS) kernel. Polynomial expansions of orders one, two and three of the PLP feature vectors were computed and concatenated to yield an expanded vector. Each of the expanded vector components was normalized to have unity variance within each speaker segment. We used these vectors as base features for the GLDS-SVM system. The base features were compensated for session variation using Nuisance Attribute Projection (NAP), with the transform being trained on the SRE04 training data and a subspace dimension of 50. We normalized each of the resulting feature components using min-max scale-and-shift across all speaker in the impostor set to normalize any dot product performed later in the SVM.

About 4000 speaker segments were taken from the SRE04 training data and were used as impostors for SVM training. We used a linear kernel C-SVM classifier (LIBSVM library). In the test phase, every trial score was normalized using T-norm with 500 cohort speakers from the NIST SRE05 training data.

The real-time runtime factors are x0.25RT and x0.12RT for the training and test phases respectively on an Intel Core2 2.4GHz CPU. Memory usage was below 1Gb.

GSV-SVM System
The GSV-SVM system uses Gaussian Mean Supervectors (GSV) as base features. A Universal Background Model (UBM) with 512 Gaussian components was trained using 3 iterations of Maximum Likelihood per Gaussian-split and 40 hours of speech training data. The UBM was adapted to the speech data of every speaker of interest using standard MAP adaptation. The mean vectors of all Gaussian components were concatenated into a supervector and each mean coefficient normalized by its standard deviation and the square root of its weight, which is equivalent to classification using the GMM Linear Kernel. These supervectors were post-processed using NAP with the same setup as in the GLDS-SVM system. SVM training and test were performed in the same way as in the GLDS-SVM system, including T-norm score normalization.

The real-time runtime factors are x0.33RT and x0.19RT for the training and test phases respectively on an Intel Core2 2.4GHz CPU. Memory usage was below 1Gb.

Submission
We submitted two systems. The primary submission consisted of the GLDS-SVM system alone. A contrastive system was submitted as the score average of the GLDS-SVM and GSV-SVM scores with 0.75 and 0.25 weights respectively.