Researchers have developed a system called Aslema for a shared task at the NADI 2026 conference. The system consists of two subtasks: intent recognition and slot filling. It uses four large language models (LLMs) in a zero-shot setting, comparing them to fine-tuned models. The results show that fine-tuning outperforms zero-shot inference. To improve performance, the researchers also explored synthetic data augmentation by generating culturally grounded Tunisian Derja utterances using an LLM and voice cloning to create synthetic speech. Incorporating this synthetic data improves performance on both tasks. The final submitted system achieves high accuracy in intent recognition and slot filling. --- Why it matters: This matters because it shows that fine-tuning large language models can improve their performance, especially when combined with synthetic data augmentation. This is relevant for researchers working on natural language processing (NLP) and speech recognition tasks. Source: https://arxiv.org/abs/2608.18689