This repo contains code for the paper: EMVD dataset: a dataset of extreme vocal distortion techniques used in heavy metal [1]. It gives instruction to reproduce the baseline experiments shown in the paper (sections 2 and 3), as well as example codes for using the EMVD dataset metadata correctly (section 4).
The codebase is developed with Python 3.9.15. To install the required dependencies, run the following command:
pip install -r requirements.txt
Run the following command to download the EMVD dataset from zenodo:
python3 download_dataset.py
First, create a Mel dataset stored in a h5 file using this command:
python3 create_dataset.py -dataset kfold
Run the following commands to train the model, store the predictions, and calculate the metrics that are used in the paper:
python3 experiment.py -exp t_multiclass -step train
python3 experiment.py -exp t_multiclass -step eval
python3 experiment.py -exp t_multiclass -step metric
You can run the previous lines with -exp t_binary to train on binary classification instead of multi-class classification.
After running the metric step, results are plotted in the terminal, and the confusion matrix for the multi-class classification is saved in the results folder.
Note that if you want to do the experiment again using another seed for the kfold cross-validation or another number of folds, you can run the following command that will save the split as new_split_kfolds.csv in the main directory:
python3 create_split_k_fold.py -seed 0 -n_splits 4
You'll then just have to replace the old file with the new one in the folder EMVD.
To replicate the figure that shows the total duration and the number of singers for each technique in the dataset, run:
python3 plot_dataset_analysis.py
To replicate the spectrograms shown in the paper, run:
python3 plot_spectrograms.py
There are two different metadata files: metadata_files.csv which contains metadata for each audio file recorded, and metadata_singers.csv which contains metadata for each singer. The two files can be joined on the column singer_id. An example of how to join the two metadata files with pandas can be found in the jupyter notebook load_metadata.ipynb.
[1] link to our paper