Compact Amharic OCR: Evaluation on Unseen Writers in the Fidel Dataset

Amharic optical character recognition (OCR) remains challenging because handwriting varies considerably across writers. Large recognition systems may also be difficult to deploy where computing resources are limited. We present a compact recognizer for segmented Amharic text lines that combines a convolutional image encoder, a lightweight visual sequence model (LightSVTR), and Connectionist Temporal Classification (CTC) decoding, with an auxiliary attention-based decoder used only during training. We evaluate the model on two held-out test sets: a custom test whose handwritten writers do not appear in the training or validation data, and the official Fidel test. The selected 24.46M-parameter model achieves 6.99% macro character error rate (CER) on the custom test and 7.90% on the official test. On the handwritten portions, macro CER increases to 18.73% and 17.95%, respectively, confirming that handwriting remains the main challenge. The model processes 221.53 images per second at batch size 16 on one NVIDIA V100. Together, these results establish a reproducible accuracy–efficiency baseline for Amharic line recognition and provide a foundation for improving recognition across diverse Amharic writing styles.