Neural style transfer applied to images has received considerable interest and has triggered many research activities aiming to use the underlying strategies for manipulation of music or sound. While the many fundamental differences between sounds an images limit the usefulness of a direct translation recent research in the Analysis/Synthesis team has demonstrated that a rather similar approach to what is used to manipulate painting style in images allows for quasi transparent analysis/resynthesis of sound textures. Instead of working on 2D images in the case of sound textures the convolutional networks work on the complex STFT.
The presentation will introduce the Xtextures command line software that is available in the Forum ans that allows using these techniques not only for resynthesis of textures but also in a more creative way for texturization of arbitrary sounds