This article tackles the question of stress assignment for lexical blends. The analysis is based on the study of the lexical stress pattern of 268 English and 253 Serbian blends. The combination of the stress pattern homology rule and the last stressed nucleus rule leads to correct overall predictions in 81% of cases in English and 96% of cases in Serbian, and up to respectively 89% and 98% for the prototypical (and largest) formal subtype of lexical blend, coined by hind clipping of the first source lexeme and fore clipping of the second one.