SongFormer: Scaling Music Structure Analysis with Heterogeneous Supervision
Python License arXiv GitHub HuggingFace Space HuggingFace Model Dataset SongFormDB Dataset SongFormBench Discord ASLP

Running on CPU hardware: analysis takes a few minutes per song. On ZeroGPU hardware each file would consume daily GPU quota (anonymous 2 min, free 5 min, PRO 40 min).

📌 Examples
Click to load example

Detected Music Segments