Calibrating neural networks for synthetic speech detection: A likelihood-ratio-based approach

In this paper, we introduce a calibration procedure designed to convert the uncalibrated output scores of neural networks for synthetic speech detection into calibrated and interpretable likelihood ratios. This procedure is based on the assumption that the networks subject to calibration are deterministic and have undergone training until they reached convergence. Provided these conditions are satisfied, it is then possible to transform their output values into likelihood ratios using a minimal set of validation and calibration data, eliminating the need for retraining the models. We successfully tested the entire workflow on a state-of-the-art network example, demonstrating not only its effectiveness in calibration but also its ability to enhance fault tolerance against inadequate inputs.