Some spoofing attacks shape speech-deepfake detectors more than others, and a new training plan can reduce that bias | arXiv News