Singer-identification-in-artist20

The source code of "Addressing the confounds of accompaniments in singer identification"

• arxiv: https://arxiv.org/abs/2002.06817

Dependencies

Requires following packages:

• python 3.6
• pytorch 1.3
• crepe 0.0.10
• librosa 0.7.1
• dill 0.3.1.1
• tqdm
• h5py
• sklearn

Usage

extract_fea.py

Extracting melspectrograms of artist20

1. Origin: the original artist20, containing both vocals and accompaniments. art_dir: path to artist20

2. Vocal: the vocal-only artist20, separated by open_unmix.art_dir: path to pure vocals of artist20 (the folder structure should follow the artist20's)

3. Accompaniment: the accompaniment-only artist20 (bass+drums+other), separated by open_unmix. art_dir: path to pure accompaniments of artis20 (the folder structure should follow the artist20's )

extract_melody.py

extract the melody of vocals using crepe

train_CRNN.py

usage: train_CRNN.py [-h] [-class CLASSES_NUM] [-gid GPU_INDEX]
                     [-bs BATCH_SIZE] [-lr LEARN_RATE] [-val VAL_NUM]  
                     [-stop STOP_NUM] [-rs RANDOM_STATE] [--origin] [--vocal]
                     [--remix] [--all] [--CRNNx2] [--debug]

optional arguments:
  -class, classes number (default:20)
  -gid, gpu index (default:0)
  -bs, batch size (default:100)
  -lr, learn rate (default:0.0001)
  -val, valid per epoch (default:1)
  -stop, early stop (default:20)
  -rs random state (default:0)
  --origin, use original audio to training
  --vocal, use separated vocal audio to training
  --remix, use remix audio to training
  --all, use all of the above data to training
  --CRNNx2, use CRNNx2 model to training
  --debug, debug mode

predict_on_audio.py

python predict_on_audio.py your_song_path