Multilingual Semantic Retrieval for Apple Music Search

Apple Music serves listeners across 150+ storefronts in dozens of languages, with a catalog growing by hundreds of thousands of new tracks daily. At this scale, search recall on misspelled, transliterated, and cross-lingual queries becomes a dominant driver of session quality, particularly for tail queries that account for the majority of unique queries. The existing token-based index struggled with these hard queries, creating a clear tension between serving popular head queries reliably and retrieving relevant results for the long tail of user intent.

The team built a multilingual semantic retrieval system centered on a 305M-parameter Siamese bi-encoder fine-tuned from GTE-multilingual-base with curriculum-scheduled multi-objective training. They integrated it into the search stack via a hybrid architecture that blends dense nearest-neighbor results with the existing token-based index using quantile distribution matching, which avoids retraining downstream rankers. Offline, the model achieved a 69% relative improvement in Hit@10 over the base model. In a worldwide online A/B test, the system delivered a 2.28% relative conversion-rate lift overall and an 86% reduction in the no-result rate, with gains across every storefront and no regressions.

For builders of large-scale search systems, the key takeaway is that semantic retrieval can dramatically improve recall on hard queries without disturbing well-served popular ones. The improvement is concentrated where it is needed most: tail queries saw a 7.93% relative CR lift, compared with 0.89% for mid-frequency and 0.14% for head queries. The hybrid architecture’s ability to blend dense and sparse signals without retraining downstream components makes this a practical, deployable approach. This work demonstrates that careful multi-objective training and curriculum scheduling can produce a production-grade multilingual retrieval model that scales across languages and storefronts.