Koshur Diacritizer: A Byte-Level Sequence-to-Sequence Model for Kashmiri Diacritic Restoration

Researchers have introduced Koshur Diacritizer, a byte-level sequence-to-sequence model designed to restore missing diacritic marks in Kashmiri digital text, a persistent source of ambiguity for the Indo-Aryan language written in a modified Perso-Arabic script [1]. The framework, detailed in a paper submitted on 14 June 2026, is built on a ByT5-small architecture and combines script-aware normalization, alignment validation, and skeleton-preserving inference to ensure that the original base-letter sequence remains intact during restoration [1][2]. The team, led by Haq Nawaz Malik, released a publicly available dataset of 23.7k aligned undiacritized and diacritized Kashmiri sentence pairs to support the task [1][2]. On a held-out test set, the model recorded a diacritic error rate (DERm) of 0.2012 and a word error rate (WER) of 0.2159 [1][2]. A separate evaluation conducted by a native Kashmiri linguistic expert yielded a mean accuracy of 77.5% [1][2]. The dataset, model weights, and source code have been made publicly available to establish a reproducible baseline for Kashmiri diacritic restoration [1][2]. The work addresses a gap in low-resource language processing, where the frequent omission of diacritics in digital communication hampers downstream natural language processing applications [1][2]. The release of aligned corpora for underrepresented languages remains a critical step for building practical text-normalization tools, and the authors position the Koshur Diacritizer as a foundation for future research in this area [1][2].