Multilingual Semantic Retrieval for Apple Music Search

Apple Music serves users across more than 150 storefronts in numerous languages, operating a massive catalog to which hundreds of thousands of new songs are added every day. In this environment, improving search recall for typos, transliteration, and cross-language queries is a key challenge for user experience.

To address this, the team developed a 305M-parameter Siamese bi-encoder model based on GTE-multilingual-base, applying multi-objective training with curriculum scheduling. This model was integrated into the search stack through a hybrid retrieval architecture that combines the existing token-based index with quantile distribution matching, without requiring retraining of the existing ranker.

Performance measurements showed a 69% relative improvement in the Hit@10 metric compared to the existing model in offline testing. The results of an actual global A/B test are as follows:

• Overall conversion rate (CR) increased by 2.28%
• No-result rate decreased by 86%
• Tail queries (low-frequency queries) achieved a high CR increase of 7.93%

Notably, this update showed the greatest effect on hard-to-search tail queries, meaning it dramatically improved search quality for complex queries without degrading performance on existing popular search terms.