What it does

This repository is a hand-maintained directory of demo websites for automatic music generation research. It sorts projects into three buckets: song generation (lyrics-to-song with vocals and backing), text-to-music, and text-to-audio. Each entry links to the paper, notes the architecture (transformer, diffusion, flow matching, or the occasional mamba), and provides a working URL.

The interesting bit

The value isn’t in the code—there is none—but in the curation. The maintainer tracks a field that moves fast enough to make arXiv feel sluggish, and the taxonomy reveals where the research heat is: diffusion still dominates, flow matching is rising, and someone actually built a music generator with Mamba.

Key highlights

• 100+ entries spanning roughly 2023–2026 research, with direct links to demos and papers
• Architecture labels on every entry, making it trivial to spot methodological trends (diffusion vs. transformer vs. flow)
• Three clear categories: song generation (the hard problem: vocals + structure), text-to-music (instrumental), text-to-audio (sound effects and foley)
• Working URLs mostly verified: a few N/A or broken links noted honestly (ERNIE-Music, MambaFoley)
• Covers major labs: Stability AI, Google Research, Microsoft, Meta, plus academic groups

Caveats

• No code, no benchmarks, no evaluation—just links; this is pure reference material
• Some demo links rot; the list appears updated but not exhaustively audited
• A few entries use rebrand.ly or notion.site links that may not age well

Verdict

Grab this if you’re building in AI audio and need to know what already exists before you reinvent MusicGen. Skip it if you want implementations; most entries link to demo pages, not repos.