How Spotify Rebuilt Its Recommendation Engine for Podcasts

Episode notes

In episode 103 of The CTO Podcast, Lucas and Luna explore the architectural overhaul behind Spotify's podcast recommendation system. They break down how Spotify moved from a music-first collaborative filtering model to a neural network that understands spoken content, processes 500 million daily listening events, and serves 100 million podcast listeners. The conversation covers the shift from sparse audio features to NLP embeddings, the decision to abandon user-item matrices for two-tower neural nets, and how the team solved cold-start problems for new shows. Lucas explains why podcast recommendations require fundamentally different engineering than music, and Luna challenges whether personalization trades off against serendipity. A focused look at one of the most complex recommender systems in the world. #Spotify #PodcastRecommendation #RecommendationEngine #MachineLearning #NeuralNetworks #NLP #AudioEmbeddings #CollaborativeFiltering #ColdStartProblem #Personalization #TwoTowerModel #Technology #Business #FexingoBusiness #BusinessPodcast #CTOPodcast #TechLeadership #EngineeringArchitecture Keep every episode free: buymeacoffee.com/fexingo

Related episodes across the Index

Other episodes covering the same guests and topics, from across The B2B Podcast Index.

• Episode 40: The Economic Reality of AI: Friction, Talent, and the Future of the FirmHigh Signal · on natural language processing97 / 100
• CMMC Is an HR Problem, Not an Enclave Problem - Here's the ProofSecure Talk Podcast · on natural language processing95 / 100
• 62: How an ALS Diagnosis Inspired an AI Companion (with Kirstin Maurer)AI Product Leader · on natural language processing80 / 100
• What Customers DO Tells You More Than What They SAYColorado Tech People · on natural language processing80 / 100
• The CXLive! Episode 98: The Insight Flywheel: Making EBC Conversations Count in The Age of AI with Dmitry RisukhinThe CX Live! · on natural language processing78 / 100
• Timothy Smith on AI, Data, and Drug Discovery at TakedaPharma Sessions · on natural language processing78 / 100

• How Datadog Scaled Engineering Without Burning Out82 / 100
• How Slack Scaled Engineering Without Adding Headcount
• How Atlassian Tamed Technical Debt With Architecture Contracts
• How Atlassian Tamed Technical Debt With Architecture Contracts
• How GitHub Tamed Merge Conflict Chaos