Hidden Echoes Survive Training in Audio To Audio Generative Instrument Models

In our work, though, we do not seek to influence As generative techniques pervade the audio domain, the behavior of the model so drastically, but rather to there has been increasing interest in tracing back through "tag" the data in such a way that the model reproduces these complicated models to understand how they draw the tag, similarly to how [10] watermark their training on their training data to synthesize new examples, both data for a diffusion image model. We are also inspired to ensure that they use properly licensed data and also to by the recent lawsuit by Getty Images against Stable elucidate their black box behavior. In this paper, we show Diffusion when it was discovered that the latter would that if imperceptible echoes are hidden in the training often reproduce the former's watermarks in its output data, a wide variety of audio to audio architectures (differentiable [29]. We would like to do something similar with audio, digital signal processing (DDSP), Realtime but to keep it imperceptible. Audio Variational autoEncoder (RAVE), and "Dance All of the above approaches use neural networks to Diffusion") will reproduce these echoes in their outputs.