AES Show 2024 NY

MOTIVATION: Generative AI for audio is quickly proliferating in both the commercial and open source software communities, However, there is still no technical consensus on how novel the generated audio data is, and what (if any) characteristics from the training data are commonly replicated. This workshop will explore existing technical approaches for how to detect memorized audio training data, how often it happens, and what characteristics of the audio are memorized. Additionally, we plan to share recent research on how audio similarity algorithms can be used to attribute audio samples produced by a generative model to specific audio samples in the training set.

FORMAT: this will be in the form of a panel discussion with three parts to it (listed below) – the purpose is to inform the AES community about the technical feasibility of detecting memorization and attributing training data in audio generative models, give audio examples of the results, an outlook for future developments, and solicit feedback from the audience.

• Memorization and generalization in deep learning
• Searching for memorized training data in audio generative models
• Training data attribution using audio similarity measures