music-detection-enhanced

Music Detection enhanced analysis models

Models used sequentially by https://github.com/Hikari-Tsai/music-detection: single standard HTDemucs vocal separation, release separator, then GAME Large v1.0.3 vocal note estimation. No additional training or quantization.

• demucs/htdemucs.onnx: community split-graph FP32 export by Ghilda, sourced from https://huggingface.co/Ghilda/htdemucs-onnx at revision 850cd89461d0817276337061c29e6abceb86d75f. Publisher states weights come from official Meta htdemucs. MIT, Meta Platforms and Ghilda; notices in demucs/. This is not an official Meta ONNX release. Needs external STFT/iSTFT, stereo 44.1 kHz, fixed 343980 samples. Source order drums, bass, other, vocals.
• game-large/: unmodified official GAME v1.0.3 Large ONNX files from https://github.com/openvpi/GAME/releases/tag/v1.0.3. Model weights are CC BY-NC-SA 4.0; software is separately MIT. Noncommercial model use only, with attribution and share-alike conditions. See bundled license. The host application's MIT license does not replace model licenses.

HTDemucs paper: Simon Rouard, Francisco Massa and Alexandre Défossez, Hybrid Transformers for Music Source Separation, ICASSP 2023, https://arxiv.org/abs/2211.08553. GAME: attribution and upstream model card https://github.com/openvpi/GAME.

Model predictions are estimates. Separated vocals may retain backing vocals or instrumental bleed. This mirror is not a quality or pitch-accuracy benchmark.