Abstract: An audio signal is segmented into a plurality of audio signal segments. A plurality of modified audio signal segments are generated based on processing data from the plurality of audio signal segments using a trained machine learning model. The plurality of modified audio signal segments are indistinguishable from the plurality of audio signal segments to the average human listener. A reconstructed audio signal corresponding to the audio signal is generated by combining the plurality of modified audio signal segments. The reconstructed audio signal is indistinguishable from the audio signal to the average human listener but, when used to train a generative machine learning model, constrains an ability of the trained generative machine learning model to generate new audio signals similar to the audio signal.