We present our submission to the KSAA-2026 Shared Task (Subtask 2): Automatic Diacritization of Speech Dictation. Building upon the CATT-Whisper multimodal architecture, which fuses representations from a pre-trained CATT text encoder and the Whisper speech encoder, we fine-tune the model end-to-end on the official shared task training data. To further enhance performance on speech-dictated Arabic text, we apply careful post-processing to the model outputs. Our best submission achieves a Diacritic Error Rate (DER) of 7.04, a Word Error Rate (WER) of 24.39, and a Sentence Error Rate (SER) of 71.65 on the hidden test set, securing 2nd place in the competition. These results demonstrate the effectiveness of adapting a strong multimodal baseline to the speech-aware diacritization setting and highlight the value of task-specific fine-tuning and output refinement for bridging the gap between spoken transcripts and fully diacritized Arabic text.

This paper describes the Rosetta system for Subtask 1 (Context-Aware English-to-Dialectal Arabic Dialogue Translation) of the AlexandriaX shared task, participating in both constrained and unconstrained tracks. The approach fine-tunes a LoRA adapter on NileChat-3B using structured system/user prompts that condition generation on dialect and dialogue context. For the unconstrained track, the adapter is additionally pretrained on MADAR and PADIC. Rosetta ranked 4th in the constrained track (spBLEU 26.10) and 5th in the unconstrained track (spBLEU 25.09). The experimental results demonstrate that external pretraining helps only two of thirteen dialects while slightly hurting overall performance, suggesting negative transfer.