aim-qmul/diffvox

Differentiable Vocal Effects Presets Dataset

This dataset is associated with DiffVox, a novel and interpretable model designed for matching vocal effects in music production. DiffVox integrates parametric equalisation, dynamic range control, delay, and reverb using efficient differentiable implementations, enabling gradient-based optimisation for parameter estimation. The vocal presets contained within this dataset were retrieved from two collections: 70 tracks from MedleyDB and 365 tracks from a private collection. This work lays the foundation for future research in vocal effects modelling and automatic mixing.

This dataset is a clone of a subset of the diffvox repository, containing a collection of vocal effect presets derived from a proprietary multitrack dataset. It contains the minimal set of files for loading the presets. You can load this subset module into your Huggingface Space using the Huggingface API. For concrete examples, please refer to our demo.

Links

• Paper: DiffVox: A Differentiable Model for Capturing and Analysing Vocal Effects Distributions
• Code: https://github.com/SonyResearch/diffvox
• Demo Space: https://huggingface.co/spaces/yoyolicoris/diffvox

Dataset Structure

The preset datasets, Internal and MedleyDB, are stored in the folder presets. Please check the original repository for more details on the individual files. The modules/ contains differentiable effects implemented in PyTorch for loading the presets and applying them to audio data.

Sample Usage

For details on how we created the dataset, please refer to the original GitHub repository. Below is a quick start example demonstrating how to set up the environment and run the retrieval process on a single track:

First, install the required Python packages by cloning the original repository and installing its requirements:

git clone https://github.com/SonyResearch/diffvox.git
cd diffvox
pip install -r requirements.txt

Then, you can run the vocal effects retrieval process on a specific track. For example, to process a track from MedleyDB:

python main.py data_dir=/data/medley1/v1/Audio/AimeeNorwich_Child --dataset=medley_vocal --log_dir=~/medley_vocal_log

This command will run the retrieval process on valid vocal tracks within the specified directory and save the training logs and best checkpoints to ~/medley_vocal_log. Refer to the original repository's main.py and cfg/config.yaml for configuration details.