Prompt-Based Music Discovery: A Prototype Using Source Separation And LLMs

We present the prototype of a music recommendation system enabling users to make specific requests using natural language prompts, such as "recommend hip-hop songs without drums but with guitar." Our approach combines audio source separation with large language model (LLM) query processing to bridge the gap between user intent and music discovery. The system uses Demucs 6-source separation to extract instrument presence information from 604 songs across 8 genres, stores metadata in a structured database, and employs Groq’s Llama-3.3-70B-Versatile model to convert natural language queries into SQL statements. While current capabilities focus on instrument presence detection, the system architecture supports future expansion to other MIR tasks. This prototype demonstrates the integration of music source separation (MSS) in recommendation systems.