MAAP LAB 🎵 — Music AI & Audio Research

“Why are there so few labs in Korea dedicated to Music AI? We built one.”

Mission

Advance the foundations of Music AI through practical research in tagging, generation, and dataset-centric methods — then share our results openly with the community. ✨

Open Science

We aim to publish at top venues (e.g., ICASSP, ISMIR, AAAI) and release code, models, and datasets whenever possible. 📢

Latest News 🗞️

• ✅ NeurIPS Workshop 2025:
  • Accepted 🎉 AIBA: Attention-based Instrument Band Alignment for Text-to-Audio Diffusion
• ✍️ ICASSP 2026 submission:
  • Resubmitted and extended version of Jamendo-QA
  • Preprint available on arXiv (2509.15662)
• 🧰 GPU resources via university support: NVIDIA A100, A6000, RTX 4090 ⚙️

Our Activities 🎯

Project 1 — Music Tagging (Completed) 🏷️

• Built a tagging & augmentation pipeline with CLAP, Beam Search, Stable Audio
• Focus: dataset augmentation/creation for future work
• Targets: short-term word generation → long-term sentence generation with LLMs
• Outcome: 2 NeurIPS Workshop submissions (1 accepted, 1 rejected → resubmitted to ICASSP)
  _{AIBA — Accepted ✅ · Jamendo-QA — Arxiv 🔗}

Project 2 — Efficient Music Generation (In Progress) 🎶

• Exploring Diffusion & DiT (e.g., Flux)
• LoRA/Adapters to avoid full fine-tuning
• Goal: robust generation for data-scarce genres/instruments/domains
• Roadmap: dataset curation → baseline reproduction → adapter experiments → ablations → release

Publications & Submissions 📚

• AIBA: Attention-based Instrument Band Alignment for Text-to-Audio Diffusion — Accepted at NeurIPS Workshop 2025
• Jamendo-QA: A Large-Scale Music Question Answering Dataset — Submitted to ICASSP 2026 · Preprint on arXiv

Get Involved 🤝

Interested in collaborating on Music AI? We welcome discussions on datasets, evaluation, and model design.
Contact: [email protected] ✉️

_{© 2025 MAAP LAB • Built with ❤️ for music & AI.}