Self-Supervised VQ-VAE for One-Shot Music Style Transfer

This is the code repository for the ICASSP 2021 paper Self-Supervised VQ-VAE for One-Shot Music Style Transfer by Ondej Cfka, Alexey Ozerov, Umut imekli, and Gal Richard.

Copyright 2020 InterDigital R&D and Tlcom Paris.

Links

🔬 Paper preprint [pdf] 🎵 Supplementary website with audio examples 🎤 Demo notebook 🧠 Trained model parameters (212 MB)

Contents

• src the main codebase (the ss-vq-vae package); install with pip install ./src; usage details below
• data Jupyter notebooks for data preparation (details below)
• experiments model configuration, evaluation, and other experimental stuff

Setup

pip install -r requirements.txt
pip install ./src

Usage

To train the model, go to experiments, then run:

python -m ss_vq_vae.models.vqvae_oneshot --logdir=model train

This is assuming the training data is prepared (see below).

To run the trained model on a dataset, substitute run for train and specify the input and output paths as arguments (use run --help for more information). Alternatively, see the colab_demo.ipynb notebook for how to run the model from Python code.

Datasets

Each dataset used in the paper has a corresponding directory in data, containing a Jupyter notebook called prepare.ipynb for preparing the dataset:

• the entire training and validation dataset: data/comb; combined from LMD and RT (see below)
• Lakh MIDI Dataset (LMD), rendered as audio using SoundFonts
  • the part used as training and validation data: data/lmd/audio_train
  • the part used as the 'artificial' test set: data/lmd/audio_test
  • both require downloading the raw data and pre-processing it using data/lmd/note_seq/prepare.ipynb
  • the following SoundFonts are required (available here and here): FluidR3_GM.sf2, TimGM6mb.sf2, Arachno SoundFont - Version 1.0.sf2, Timbres Of Heaven (XGM) 3.94.sf2
• RealTracks (RT) from Band-in-a-Box UltraPAK 2018 (not freely available): data/rt
• Mixing Secrets data
  • the 'real' test set: data/mixing_secrets/test
  • the set of triplets for training the timbre metric: data/mixing_secrets/metric_train
  • both require downloading and pre-processing the data using data/mixing_secrets/download.ipynb

Acknowledgment

This work has received funding from the European Unions Horizon 2020 research and innovation programme under the Marie Skodowska-Curie grant agreement No. 765068.