Convolutional neural networks (convnets) with random initialization often fail to outperform hand-crafted filterbank baselines for audio processing tasks, due to instabilities in the energy response of the randomly initialized filters.
