Spotify has introduced audiobooks to its platform, facing challenges in personalized recommendations due to data sparsity and the need for fast, scalable models. To address these, they developed 2T-HGNN, a system combining Heterogeneous Graph Neural Networks (HGNNs) and a Two Tower (2T) model. This approach captures nuanced item relationships while ensuring low latency and complexity. The model decouples users from the HGNN graph and uses an innovative multi-link neighbor sampler, significantly reducing complexity. Empirical results show a 46% increase in new audiobook start rates and a 23% boost in streaming rates. The model also benefits existing products like podcasts. Key contributions include leveraging podcast consumption to understand audiobook preferences, a modular architecture integrating audiobook content, and a balanced sampler to address data imbalance. The system uses co-listening graphs and HGNNs to generate embeddings capturing long-range dependencies, enabling effective recommendations. The 2T model handles user preferences and interactions, ensuring scalability and real-time performance. The model is now in production, serving millions of users and improving audiobook and podcast recommendations.Spotify has introduced audiobooks to its platform, facing challenges in personalized recommendations due to data sparsity and the need for fast, scalable models. To address these, they developed 2T-HGNN, a system combining Heterogeneous Graph Neural Networks (HGNNs) and a Two Tower (2T) model. This approach captures nuanced item relationships while ensuring low latency and complexity. The model decouples users from the HGNN graph and uses an innovative multi-link neighbor sampler, significantly reducing complexity. Empirical results show a 46% increase in new audiobook start rates and a 23% boost in streaming rates. The model also benefits existing products like podcasts. Key contributions include leveraging podcast consumption to understand audiobook preferences, a modular architecture integrating audiobook content, and a balanced sampler to address data imbalance. The system uses co-listening graphs and HGNNs to generate embeddings capturing long-range dependencies, enabling effective recommendations. The 2T model handles user preferences and interactions, ensuring scalability and real-time performance. The model is now in production, serving millions of users and improving audiobook and podcast recommendations.