In this paper, we investigate a previously proposed algorithm for spoken language identification based on convolutional neural networks and convolutional recurrent neural networks. We improve the algorithm by modifying the training strategy to ensure equal class distribution and efficient memory usage. We successfully replicate previous experimental findings using a modified set of languages. Our findings confirm that both a convolutional neural network as well as convolutional recurrent neural networks are capable to learn language-specific patterns in mel spectrogram representations of speech recordings.