Towards Structured Approaches to Arbitrary Data Selection and Performance Prediction for Speaker Recognition

We developed measures relating feature vector distributions to speaker recognition (SR) performances for performance prediction and potential arbitrary data selection for SR. We examined the measures of mutual information, kurtosis, correlation, and measures pretaining to intra- and inter-speaker variability. We applied the measures on feature vectors of phones to determine which measures gave good SR, performance prediction of phones standalone and in combination. We found that mutual information had an -83.5% correlation with the Equal Error Rates (EERs) of each phone. Also, Pearson's correlation between the feature vectors of two phones had a -48.6% correalation with relative EER improvement of the score-level combination of the phones. When implemented in our new data-selection scheme (which does not require a SR system to be run), the measures allowed us to select data with 2.13% overall EER improvement (on SRE08) over data selected via a brute-force approach, at a fifth of the computational costs.