Abstract: We benchmark here several convolution kernels, particulary trying custom dilated convolutions, and show how convolutions following pitch spaces such as the Tonnetz may help the learning of musical tasks.