Adversarial Drum Synth (ADS)

TensorFlow implementation of adversarial drum synth (ADS) from the paper Adversarial Synthesis of Drum Sounds @ The 2020 DAFx Conference.

Code

Dependencies

Python

Code has been developed with Python 3.8. It should work with other versions of Python 3, but has not been tested. Moreover, we rely on several third-party libraries, listed in requirements.txt. They can be installed with

$ pip install -r requirements.txt

Usage

The majority of the code is contained within three scripts train.py, preproc.py and generate.py. The first script train.py allows to train a new model from scratch as described in the paper. The preproc.py script is used to prepare a new dataset for training. The third script generate.py enables conditional synthesises of drum sounds using a pretrained generator.

train.py arguments

| Parameter | Default | Description | | :------------------------ |:-------------:| :-------------| | --dataset | - |path to folder containing audio training data | --preproc | False |run audio preprocessing script | --n_classes | 3 |number of classes (e.g., kick, snare, cymbal) | --sr | 44100 | audio sample rate | --audio_length | 16384 | audio length in samples (must be a power of 2) | --batch_size | 64 | training batch size | --epochs | 2000 | number of cycles through the full dataset | --lr | 0.0002 | learning rate | --beta_1 | 0.5 | Adam beta 1 value | --n_d | 5 | number of discriminator updates per generator update | --z_dim | 3 | latent dimensionality | --e_dim | 50 | embedding layer dimensionality | --num_res | 5 | number of upsampling/downsampling blocks | --scale_base | 4 | upsampling/downsampling factor | --n_filters | 1024 | maximum number of filters | --adversarial_loss_mode | 'wgan' | adversarial loss mode | --gradient_penalty_mode | 'wgan-gp' | gradient penalty mode | --experiment_name | 'ADS_demo' | name of the experiment

preproc.py

When training the system on your own dataset of drum sounds, the dataset must be prepared using the preproc.py script. Create a directory structure with a sub-directories for each class containing the appropriate drum sounds. Run train.py and set the --dataset parameter to the path of your dataset and then set --preproc to True. If training on a dataset that has already been prepared with preproc.py, --dataset can be set to the path of the preprocessed dataset, and --preproc can be set to False.

generate.py arguments

| Parameter | Default | Description | | :------------------------ |:-------------:| :-------------| | --z | 0.25 -0.25 0.25 |latent vector values for generation | --condition | 0 |which class to generate (e.g., 0: kick, 1: snare, 2: cymbal) | --weights | './output/demo' |path of the pretrained generator weights to be loaded | --output_dir | './generations' |path to store generations | --sr | 44100 |audio sample rate | --name | 'generation' |name of generation

Note that a pre-trained model is available in the output/demo folder.

Supporting webpage

For more information, please visit the corresponding supporting website.

It contains the following:

• Audio examples
  • Training data
  • Generations
  • Example usage within loop-based electronic music compositions
  • Generating Drum Loops
  • Interpolation demonstration
• Supplementary figures
• A link to the DAFx 2020 paper and presentation

References

| [1] | Drysdale, J., M. Tomczak, J. Hockman, Adversarial Synthesis of Drum Sounds. Proceedings of the 23rd International Conference on Digital Audio Effects (DAFX), 2020.| | :---- | :--- |

@inproceedings{drysdale2020ads,
  title={Adversarial synthesis of drum sounds},
  author={Drysdale, Jake and Tomczak, Maciej and Hockman, Jason},
  booktitle = {Proceedings of the International Conference on Digital Audio Effects (DAFx)},
  year={2020}
}

Help

Any questions please feel free to contact me on jake.drysdale@bcu.ac.uk