Weighting Scores to Improve Speaker-Dependent Threshold Estimation in Text-Dependent Speaker Verification