[@fma_dataset]: Defferrard, Michaël, et al. "FMA: A Dataset for Music Analysis." 18th International Society for Music Information Retrieval Conference (ISMIR), 2017. arXiv:1612.01840
- classify audio by seperating channels using nussl/spleeter/demucs
https://github.com/facebookresearch/demucs/tree/main- Seperate others manually using U-net or nussl or filtering
try the app deployed at huggingface🤗. currently buggy
moedel used is `https://github.com/deezer/spleeter/tree/master/configs/5stems
There is also files from MillionSongSubset in h5 http://millionsongdataset.com
plus Classical instruments
useful other sets
https://zenodo.org/records/4599666 - slakh dataset 107Gb
https://sigsep.github.io/datasets/musdb.html#musdb18-hq-uncompressed-wav - 22.7Gb
https://github.com/lukewys/chamber-ensemble-generator#dataset-download - upto 4Tb