NADI 2026 TASK 4 AUTOMATIC SPEECH TRANSLATION

The Automatic Speech Translation subtask evaluates systems on their ability to translate spoken Arabic dialects directly into English. Given an input audio recording, participants must generate the corresponding English translation. The task covers eight Arabic dialects: Algerian, Egyptian, Jordanian, Mauritanian, Moroccan, Palestinian, Emirati, and Yemeni Arabic. Submissions are evaluated primarily using corpus-level BLEU and chrF. Per-dialect scores are also reported. For more details, visit the **official NADI 2026 page**: https://nadi.dlnlp.ai/2026/ The **dataset** is available here: https://huggingface.co/datasets/fbougares/NADI_Whitehouse_AST_2026