Introducing Natural Language Search for Podcast Episodes

This post details the technical implementation of Natural Language Search for podcast episodes at Spotify. It covers the transition from term-based search to semantic search using dense retrieval models, transformer networks (Universal Sentence Encoder CMLM), and ANN search. The architecture involves training query and episode encoders in a shared embedding space, data preparation including successful search logs and synthetic query generation, and training using a siamese network with in-batch negatives and margin loss. The integration with production uses Vespa for offline indexing of episode vectors and online query encoding and retrieval, with a first-phase ranking function for re-ranking.