A Handwritten Text Recognition Dataset for Ajami Manuscripts in Fulfulde and Hausa

We present the first ever dataset of manually segmented and transcribed Ajami manuscripts written in Fulfulde and Hausa. The term Ajami refers to modified Arabic-script orthographies in Africa. Existing handwritten text recognition (HTR) and optical character recognition (OCR) models for Arabic-script languages perform poorly on West African manuscripts due to a lack of these manuscripts representation in the models’ pre-training. This leads to models struggling to adapt to Ajami style calligraphy, being unequipped to recognize Ajami specific characters, and being unable to extract certain Arabic-script diacritics which are present in Ajami manuscripts but lacking in many manuscripts for other Arabic-script languages like Arabic and Persian. The latter poses a significant challenge to Ajami HTR. We release the following as an open-source dataset: an ALTO formatting of high-quality images of Fulfulde and Hausa manuscripts, manual segmentation (region and line), and manual transcriptions. Our HTR dataset is also the first to diplomatically transcribe newly Unicode-encoded, special Quranic recitation characters. We evaluate a suite of Arabic-script recognition models specifically for historical manuscripts and find that they produce character error rates of 65–84% when attempting to automatically transcribe our curated manuscripts. Transcriptions produced by the evaluated models are released as well.