Based on WandB and PyTorch Lightning

Modified data preprocessing to MARBLE Benchmark

Preparation

1. Download VocalSet from https://zenodo.org/record/1193957
2. Run the code

python prepare.py

Then, split files will be created.

Main program

python main.py

The detailed experiment conditions are contained in run.sh.

• APSIPA 2021: Refer to the line "OblongCNN"
• Interspeech 2022: Refer to the line "OblongDeformableCNN"
• The Best Version: Refer to the line "dcnv2 crt stft", which uses Multi-resolution spectrogram, classifier retraining, and modulated deformable convolution (DCNv2)

Reference:

@inproceedings{yamamoto22_interspeech,
  author={Yuya Yamamoto and Juhan Nam and Hiroko Terasawa},
  title={{Deformable CNN and Imbalance-Aware Feature Learning for Singing Technique Classification}},
  year=2022,
  booktitle={Proc. Interspeech 2022},
  pages={2778--2782},
  doi={10.21437/Interspeech.2022-11137}
}