We describe Abjad AI’s submission to KSAA-2026 Shared Task 2 on automatic diacritization of Arabic speech dictation. The task requires generating fully diacritized text given speech audio and an undiacritized transcript. Because text-only diacritization cannot resolve ambiguities that are recoverable from the acoustic signal, we propose conditioning a character-level encoder-only Transformer (CATT) (Alasmary et al., 2024) on speech representations. We introduce grouped speech conditioning, which downsamples speech encoder features into a small set of pooled tokens concatenated to the text input, enabling efficient fusion without architectural changes to CATT. We train with a two-phase schedule that first freezes the text encoder, then fine-tunes the full model. Our best system, using Whisper-small (Rad-ford et al., 2022) features with five grouped tokens, achieves a Diacritization Error Rate (DER) of 6.60 and a Word Error Rate (WER) of 18.66 (without case endings, including no-diacritic) on the official test set. Notably, we find that Whisper-small consistently outperforms Whisper-large-v3, suggesting that compact speech representations better suit this fusion setting. This is an extended and revised version of our previous work (Ghannam et al., 2025).

Tashkeel, or Arabic Text Diacritization (ATD), greatly enhances the comprehension of Arabic text by removing ambiguity and minimizing the risk of misinterpretations caused by its absence.It plays a crucial role in improving Arabic text processing, particularly in applications such as text-to-speech and machine translation.This paper introduces a new approach to training ATD models.First, we finetuned two transformers, encoder-only and encoder-decoder, that were initialized from a pretrained character-based BERT.Then, we applied the Noisy-Student approach to boost the performance of the best model.We evaluated our models alongside 11 commercial and open-source models using two manually labeled benchmark datasets: WikiNews and our CATT dataset.Our findings show that our top model surpasses all evaluated models by relative Diacritic Error Rates (DERs) of 30.83% and 35.21% on WikiNews and CATT, respectively, achieving state-of-the-art in ATD.In addition, we show that our model outperforms GPT-4-turbo on CATT dataset by a relative DER of 9.36%.We open-source our CATT models and benchmark dataset for the research community .