The pretrained 4stems model has the following performances when tested on the musDB test set:

The first column of this table can be reproduced using the following command (you need to download musdb first and to put it into the folder):

spleeter evaluate -p spleeter:4stems -o spleeter_mask_results --mus_dir <musdb path>

The second column can be reproduced using the following command:

spleeter evaluate -p spleeter:4stems -o spleeter_mwf_results --mus_dir <musdb path> -m

The pre-trained models were trained on a large private dataset. It is also possible to train a model with the public musDB train dataset. To do so use the configs/musdb_config.json:

spleeter train --verbose -p configs/musdb_config.json -d <musdb path>

This should provide the following results on the musDB test dataset (note that results may differ a bit due to a not perfectly deterministic data pipeline):

Once the model trained, the results of this table can be generated with:

spleeter evaluate -p configs/musdb_config.json -o musdb_trained_spleeter_results --mus_dir <musdb path> -m