Audio-Novel

Research on using Large Language Models to generate synthetic queries for improving audiobook retrievability in Spotify's search system, addressing cold-start challenges through query auto-completion and retrieval enhancement.

Spotify's innovative 2T-HGNN system combining Heterogeneous Graph Neural Networks and Two Tower models for scalable audiobook recommendations, achieving +46% start rate increase.

A zero-shot approach for generating expressive audiobooks using multimodal speaker personas and LLM-based script instructions, eliminating costly training and manual annotation.