Researchers have demonstrated the ability to “unlearn” sensitive identifying data from audio used to train machine learning models.