Empirical Calibration and Metric Differential Privacy in Language Models