The Byzantine religious tradition includes Greek Orthodox Church hymns, which significantly differ from other cultures’ religious music. Since the deep learning revolution, audio and music signal processing are often approached as computer vision problems. This work trains from scratch three different novel convolutional neural networks on a hymns dataset to perform hymns classification for mobile applications. The audio data are first transformed into Mel-spectrograms and then fed as input to the model. To study in more detail our models’ performance, two state-of-the-art (SOTA) deep learning models were trained on the same dataset. Our approach outperforms the SOTA models both in terms of accuracy and their characteristics. Additional statistical analysis was conducted to validate the results obtained.