Abstract
Southeast Asia (SEA) is a region rich in linguistic diversity and culturalvariety, with over 1,300 indigenous languages and a population of 671 millionpeople. However, prevailing AI models suffer from a significant lack ofrepresentation of texts, images, and audio datasets from SEA, compromising thequality of AI models for SEA languages. Evaluating models for SEA languages ischallenging due to the scarcity of high-quality datasets, compounded by thedominance of English training data, raising concerns about potential culturalmisrepresentation. To address these challenges, we introduce SEACrowd, acollaborative initiative that consolidates a comprehensive resource hub thatfills the resource gap by providing standardized corpora in nearly 1,000 SEAlanguages across three modalities. Through our SEACrowd benchmarks, we assessthe quality of AI models on 36 indigenous languages across 13 tasks, offeringvaluable insights into the current AI landscape in SEA. Furthermore, we proposestrategies to facilitate greater AI advancements, maximizing potential utilityand resource equity for the future of AI in SEA.