It is generally assumed that noun-noun compounds in English are stressed on the left-hand member (e.g. cóurtroom, wátchmaker). However, there is a large amount of variation in stress assignment (e.g. silk tíe, Madison Á venue, singer-sóngwriter) whose significance and sources are largely unaccounted for in the literature. This article presents a study in which three kinds of factors held to play a role in compound stress assignment are tested: argument structure, lexicalization, and semantics. The analysis of 4,353 noun-noun compounds extracted from the Boston University Radio Speech Corpus shows that there is indeed a considerable amount of variation in stress assignment. Overall, semantics turns out to have the strongest effect on compound stress assignment, whereas an approach relying on argument structure is much less successful in predicting compound stress. The article presents for the first time large-scale empirical evidence for the assumption that lexicalization has an effect on compound stress assignment. The article also makes a methodological contribution to the debate in showing that (and how) corpus-based studies using acoustic measurements can shed new light on the issue of variable compound stress.* * The first three authors wish to dedicate this article to their academic teacher, mentor, and friend Rüdiger Zimmermann, Professor and Chair of English Linguistics at the University of Marburg, on the occasion of his retirement in September 2008. Rüdiger has been a great source of inspiration for all three of us, and our research team owes much to his generosity and support. We also would like to thank two sets of Language referees and the editor Brian Joseph for their very helpful and detailed remarks on earlier versions of this article. We are also grateful for the comments we received at the following talks, colloquia, and conferences where we presented material from this article: