Create coherent mixtures from a folder with any provided stems. The package will infer the needed metadata from the audio files and use it to create mixtures. This package currently only supports .wav audio files.
This package aims to increase the diversity of instruments in mixtures used to train source-separation models.
pip install stem_mixer
Our library relies on python-soundfile for writing files, which relies on libsndfile.
For every stem, we will calculate features and save them on a .json file.
python script/metadata.py
--data_home=<path_to_stems>
--datasets="brid","musdb"if you want to manually call the extraction function to overwrite metadata:
from metadata import extraction
# define the path to your audio stem file
stem_path = "path/to/your/stem/file.wav"
# optionally, provide pre-computed metadata
metadata = {
"stem_name": stem_name,
"data_home": data_home,
"tempo": tempo,
"key": key,
"sound_class": sound_class
}
extraction(stem_path, track_metadata=metadata, overwrite=True)More examples can be found in the examples folder.
While we infer features, we support some datasets for which we can infer features from names information.
*note: if using MUSDB18, pre-pre-processing step required -->
- must save each stem with "vocals", "drums", "bass", "other" as prefix in .wav filename
- i.e. Detsky Sad - Walkie Talkie - drums.wav
- i.e. Triviul - Angelsaint - vocals.wav
- i.e. PR - Happy Daze - bass.wav
- i.e. Voelund - Comfort Lives In Belief - other.wav
first make sure pytest is installed:
pip install pytestif you would like to run our implemented tests of the metadata module, first navigate to your script folder and then do the following:
pytest ../tests/test_md.pyIf you would like to run our implemented tests of the preprocessing module, first navigate to your script folder and then do the following:
pytest ../tests/test_pre.py