ALADAN Low-Resource Arabic Dialectal Speech Translation

This technical report presents the submission of ALADAN for the low-resource track at the IWSLT 2025 conference, focusing on North Levantine Arabic to English speech translation. The report outlines the methodology employed, which builds on previous efforts from IWSLT 2024. It describes the use of a cascade ASR architecture that integrates a TDNN-F model with an upgraded Zipformer-Large variant to enhance acoustic representation. To address data sparsity, a crowd-sourced parallel corpus was created, encompassing five major Arabic dialects. The report details the improvements observed in automatic evaluation metrics across dialects when utilizing crowd-sourced data. Additionally, it discusses the performance of the system under both low-resource and high-resource scenarios, noting that while crowd-sourced data improved results in low-resource settings, it did not enhance scores in high-resource conditions. The final submission achieved a BLEU score of 20.0 on the official test set, demonstrating the effectiveness of the proposed methods.