Generating audio mixtures using deep convolutional neural networks

Deep neural networks have recently been used in a generative capacity to separate and convolve the content and style of two input images. This is done using a joint cost function during gradient descent that encodes information about style and content to iteratively calculate forward node activations. We extend this methodology to the auditory domain using sound clips converted to spectrograms using the short-time Fourier transform and discuss optimizing signal reconstruction.