A convolutional autoencoder for speckle noise reduction of numerically reconstructed holograms is presented. The neural network is trained with a generic open dataset whose images are preprocessed to emulate speckle noise. Experimental validation is provided.