azstorm/lightcut-music-analysis-models

LightCut Music Analysis Models

ONNX inference weights used by LightCut's browser-local music structure analysis. All audio decoding, source separation, feature extraction, and inference run in the user's browser.

Files

• htdemucs-core.onnx: four-stem HTDemucs source-separation core.
• harmonix-fold0.onnx through harmonix-fold7.onnx: eight-fold All-In-One music-structure-analysis ensemble.

Intended Use

These files are packaged for inference. They predict tempo, beats, downbeats, section boundaries, and section labels for music audio. The LightCut integration supports audio-bearing media between 76 and 660 seconds.

Provenance and Attribution

• All-In-One Music Structure Analyzer, Taejun Kim and Juhan Nam, WASPAA 2023. Source: https://github.com/mir-aidj/all-in-one
• Hybrid Transformer Demucs, Alexandre Defossez, 2022. Source: https://github.com/facebookresearch/demucs

The files in this repository are ONNX conversions for browser inference. They are not newly trained model weights. Please retain this attribution when redistributing derived packages.

Version

v1 is pinned by the consuming source repository to a full Hugging Face commit SHA and validates every download against its SHA-256 digest.