An Applied Computer Mathematics Approach to Transliteration: YOLOv8-Based Detection of Harah Jawoe Script

AR, Khairan (Department of Information Technology, Universitas Islam Negeri Ar-Raniry Banda Aceh, Aceh, Indonesia)
Ahmadian, Hendri (Department of Information Technology, Universitas Islam Negeri Ar-Raniry Banda Aceh, Aceh, Indonesia)
Nasihah, Aesha Durratul (Department of Physics Engineering, Universitas Islam Negeri Ar-Raniry Banda Aceh, Aceh, Indonesia)
Hamdan, Abd Mujahid (Department of Physics Engineering, Universitas Islam Negeri Ar-Raniry Banda Aceh, Aceh, Indonesia)

Zero : Jurnal Sains, Matematika, dan Terapan
Vol 9, No 1 (2025): Zero: Jurnal Sains Matematika dan Terapan

Transliteration of traditional scripts such as Harah Jawoe, a regional variant of Jawi used in Acehnese manuscripts, is essential for preserving and understanding historical texts. While previous studies have focused on isolated character recognition, this research introduces a word-level transliteration approach using the YOLOv8 convolutional neural network (CNN) architecture. A dataset of 9,000 augmented images derived from the Hikayat Aceh manuscript was used to train and evaluate four YOLOv8 variants (small, medium, normal, and big). The results show that the big model achieved the most stable and reliable performance, with a peak mAP50–95 of 72.4% and an accuracy of 99.95%. These findings highlight the model’s capability to handle the structural complexity of Harah Jawoe script at the word level. This study offers a novel contribution by integrating object detection techniques with low-resource script transliteration, with implications for AI-driven cultural preservation and the digital accessibility of Southeast Asian manuscript heritage