Unified cross-domain semantic retrieval for cold-start audiobook discovery

Recommending a new content type to users who have never tried it is a hard cold-start problem: there is no in-domain history to infer taste from. We study audiobook discovery on a large music streaming platform, where most users have never played an audiobook. Many of them, however, leave behavioral traces in other domains: podcasts, audiobooks elsewhere, videos, and related content. We turn those traces into recommendations with a text-based retrieval model. Each user is described by the titles of the items they recently consumed in each domain, each audiobook by its metadata, and one multilingual text encoder maps both into a single space, so recommendation becomes nearest-neighbor retrieval. A learned gate weights the domains per user and suppresses the ones a user lacks. The prior production model relied on collaborative co-consumption vectors, which carry no information about an item's content. Writing everything as text removes that limit and lets one dense-retrieval model serve both content matching and behavioral recommendation; it also enables title-level transfer across domains, for instance audiobook listening elsewhere informing discovery on the platform. We evaluate on a held-out worldwide cohort covering nine marketplace groups, always against a single production baseline. Music, podcast, and audiobook signals together improve Recall@10 by +27.3%, and adding a fourth signal, video, raises this to +32.8%. A per-cohort breakdown shows each signal helps the users who actually carry it: the audiobook signal is worth roughly +50% for that group, and video helps most for users whose only other signal is music. An online A/B test is underway and will appear in the camera-ready. Per company confidentiality requirements, we report relative numbers only.