This paper presents the KSAA-2026 Shared Task on Arabic Speech Dictation with Automatic Diacritization, addressing a persistent challenge in Arabic NLP. The task focuses on transforming speech transcripts into fully diacritized Arabic text by leveraging both the speech signal and its undiacritized transcript. Unlike conventional ASR tasks that focus on transcription, this task integrates acoustic and textual information to improve diacritization accuracy. The shared task consists of two subtasks: (1) Data Contribution, where participants recorded and reviewed speech data through the VoiceWall platform, resulting in 2,160 recordings, and (2) Diacritization, where 5 teams developed systems that generate fully diacritized text from speech and undiacritized transcripts. The dataset includes approximately 5 hours of Modern Standard Arabic (MSA) and multi-dialectal speech with fully diacritized references. Experimental results show that several participant systems outperform the provided baselines, and that incorporating speech information and fine-tuning improves performance compared to text-only approaches. KSAA-2026 shared task establishes a benchmark for multimodal Arabic diacritization and supports the development of robust systems for applications in education, accessibility, and speech-driven text generation.

The linguistic inclusivity of Large Language Models (LLMs) such as ChatGPT, Gemni, JAIS, and AceGPT has not been sufficiently explored, particularly in their handling of low-resource languages like Arabic compared to English. While these models have shown impressive performance across various tasks, their effectiveness in Arabic remains under-examined. Punctuation, critical for sentence structure and comprehension in tasks like speech analysis, synthesis, and machine translation, requires precise prediction. This paper assesses seven LLMs: GPT4-o, Gemni1.5, JAIS, AceGPT, SILMA, ALLaM, and CommandR+ for Arabic punctuation prediction. Additionally, the performance of fine-tuned AraBERT is compared with these models in zero-shot and few-shot settings using a proposed Arabic punctuation prediction corpus of 10,044 sentences. The experiments demonstrate that while AraBERT performs well for specific punctuation marks, LLMs show significant promise in zero-shot learning, with further improvements in few-shot scenarios. These findings highlight the potential of LLMs to enhance the automation and accuracy of Arabic text processing.