Compact Amharic OCR: Efficient Recognition and Error Analysis of Ethiopic Handwriting

Optical character recognition (OCR) for Amharic remains particularly challenging for handwritten text, where handwriting variation interacts with subtle visual differences among related Ethiopic syllabic forms. We investigate whether a compact recognizer can provide accurate Amharic OCR across handwritten, typed, and synthetic text while identifying the factors that continue to limit handwritten performance. We develop a 24.48M-parameter line recognizer informed by lightweight OCR architectures. A compact convolutional backbone extracts visual features, height compression forms a horizontal sequence, and a LightSVTR-style module models contextual dependencies along the text line. A Connectionist Temporal Classification (CTC) head predicts 354 visible symbols plus the CTC blank. During training, an auxiliary NRTR decoder provides additional contextual sequence supervision but is removed entirely at inference, preserving a compact deployment path. Training uses 112,716 Fidel line images with a modality curriculum that progressively shifts emphasis from synthetic to typed and handwritten samples before balanced consolidation. On the 18,011-line official Fidel test set, the model achieves 8.24% macro character error rate (CER) and 6.92% micro CER. Macro CER varies substantially by modality: 3.18% on synthetic text, 4.05% on typed text, and 17.41% on handwritten text. For checkpoint selection, 41 recorded handwritten writers are withheld from training. On the official handwritten test set, performance is nearly identical for training-seen and training-held-out writers (17.42% vs. 17.48% macro CER), suggesting that writer identity alone does not explain the handwritten performance gap. Across 6,035 handwritten test lines, we observe 58,222 character-level errors: 66.20% substitutions, 27.75% deletions, and 6.05% insertions. Frequent substitutions involve visually related Ethiopic forms including ሰ/ስ, ላ/ለ, አ/እ, ደ/ዳ, and ባ/በ. These results indicate that fine-grained discrimination among similar handwritten Ethiopic glyphs is a more important remaining bottleneck than writer identity or spacing alone. Overall, the study demonstrates strong Amharic line recognition with a compact inference architecture while identifying a concrete target for future improvement: representations that better separate subtle handwritten character distinctions.