Apple Music offers listeners 150+ storefronts in multiple languages, with a catalog that grows by hundreds of thousands of new tracks every day. At this scale, search recall on misspelled, untranslated, and multilingual queries becomes a dominant driver of session quality, especially on tail queries that result in many unique queries. We present a multilingual semantic retrieval system built on the 305M-parameter Siamese fine-tuned from the GTE-multilingual-base with multi-objective training programmed into the curriculum. The model is integrated into the search stack with a hybrid retrieval structure that combines nearest neighbor results with an index based on existing tokens using quantile distribution matching, allowing use without retraining downstream estimators. Offline, the model achieves a 69% relative improvement in Hit@10 over GTE-multilingual-base. In a global online A/B test, the system delivered a relative conversion rate (CR) of 2.28% overall, an 86% drop in bounce rate, and a return on the entire storefront with no regressions noted. The improvement is focused where it is most needed: the tail questions see a relative CR increase of 7.93%, compared to 0.89% of the middle frequency questions and 0.14% of the main questions—proof that the semantic retrieval improves the recall of the difficult questions without affecting the well-presented popular ones. As far as we know, this is one of the biggest search quality improvements implemented on the platform.
