As generative techniques pervade the audio domain, there has been increasing interest in tracing back through these complicated models to understand how they draw on their training data to synthesize new examples, both to ensure that they use properly licensed data and also to elucidate their black box behavior. In this paper, we show that if imperceptible echoes are hidden in the training data, a wide variety of audio to audio architectures (differentiable digital signal processing (DDSP), Realtime Audio Variational autoEncoder (RAVE), and ``Dance Diffusion'') will reproduce these echoes in their outputs. Hiding a single echo is particularly robust across all architectures, but we also show promising results hiding longer time spread echo patterns for an increased information capacity. We conclude by showing that echoes make their way into fine tuned models, that they survive mixing/demixing, and that they survive pitch shift augmentation during training. Hence, this simple, classical idea in watermarking shows significant promise for tagging generative audio models.
Below are a few examples of style transfer from a few of the models we trained. Drums models are trained on the groove dataset, guitar models are trained on guitarset, and vocal models are trained on vocalset. For models with echoes/echo patterns, that same echo pattern was embedded in each file of the training set before training. See our paper for statistics on much more extensive experiments.
Notes:
A 30 second clip from Prince - Loring Park Sessions '77
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
Clean |
50 Echo |
75 Echo |
100 Echo |
