Music generation personalizes albums using user preference profiles
PAG-Agent: Zero-Shot Personalized Music Album Generation via Aesthetic Profile Distillation
MultimediaSound
Summary
Creating music that matches someone's tastes is hard because preferences are personal and varied. The authors introduce PAG-Agent, a system that learns a listener's unique style from their favorite songs and other info without needing extra training. This personal taste profile is then used to instruct music generators to create albums suited to that listener. They also develop new ways to measure how well the generated music fits individual preferences, showing improvements and correlations with people’s own likes.
What this means in practice
- •For music streaming platforms: Create personalized music album recommendations by generating songs tuned to each listener’s unique preferences extracted from their listening history.$Commercial implications: Allows streaming services to offer uniquely generated albums that match user tastes for enhanced engagement and subscription value.
- •For mobile app developers: Develop apps that generate custom music albums on demand for users without requiring manual preference input or additional training data.
Authors
Yifeng Mao, Ziye Wang, Yutong Shen, Feiran Liu, Wenjuan Han
Abstract
Large Music Models (LMMs) can synthesize high-fidelity audio and support increasingly flexible prompt control, yet modeling the subjective preferences of an individual listener remains underexplored. We present PAG-Agent, a zero-shot multimodal framework that uses an explicit Personal Aesthetic Profile to bridge user preference inference and music generation control. A multimodal LLM distills favorite-song audio and contextual user information into this structured representation, which a planning LLM translates into generation instructions for existing music synthesizers. We further introduce PA-Eval, which evaluates personalization through the Aesthetic Consistency Score (ACS) for target-user alignment and the Aesthetic Specificity Index (ASI) for cross-user specificity. Experiments with Personal-100 on multiple MLLM and music-generation backbones show consistent improvements over generic and direct conditioning. A 50-participant follow-up study further shows a correlation $r=0.64$ between ACS and subjective preference.