Reviving an Endangered Script: Optical Character Recognition for Syriac

The results show that while CNNs perform well on single-character recognition, Transformer-based models outperform CRNN-LSTMs on the more complex word-level task.