A Variational Y-Autoencoder for Disentangling Gesture and Material of Interaction Sounds

Appropriate sound effects are an important aspect of immersive virtual experiences. Particularly in mixed reality scenarios it may be desirable to change the acoustic properties of a naturally occurring interaction sound (e.g., the sound of a metal spoon scraping a wooden bowl) to a sound matching the characteristics of the corresponding interaction in the virtual environment (e.g., using wooden tools in a porcelain bowl). In this paper, we adapt the concept of a Y-Autoencoder (YAE) to the domain of sound effect analysis and synthesis. The YAE model makes it possible to disentangle the gesture and material properties of sound effects with a weakly supervised training strategy where only an identifier label for the material in each training example is given. We show that such a model makes it possible to resynthesize sound effects after exchanging the material label of an encoded example and obtain perceptually meaningful synthesis results with relatively low computational effort. By introducing a variational regularization for the encoded gesture, as well as an adversarial loss, we can further use the model to generate new and varying sound effects with the material characteristics of the training data, while the analyzed audio signal can originate from interactions with unknown materials.