serdjent/disentangled-vae

Polyphonic Music Style Transfer

This repository contains the code for a Master's dissertation project focused on transferring the musical style between polyphonic composers (e.g., making a piece by Bach sound as if it were composed by Chopin).

Core Models Implemented

• —Conditional Variational Autoencoder (CVAE): A generative model that learns to disentangle musical content from style.
• —Cycle-consistent Transformer: A GAN-style architecture with two Transformer models that learn to translate between two musical styles in a cycle.
• —Baseline Transformer LM: A simple language model used to verify the data representation and for basic music generation.
• —Style Classifier: An LSTM-based model trained to identify a composer's style, used as a loss component for training the CVAE.

Tech Stack

• —Python 3.10+
• —PyTorch & PyTorch Lightning
• —pretty_midi & music21 (for MIDI processing)
• —librosa (for audio analysis)
• —pandas & numpy

Setup

1. 1.Clone the repository:

    git clone [https://github.com/djentlehead/polyphonic-style-transfer.git]
    cd polyphonic-style-transfer

1. 1.Create a virtual environment and install dependencies:

    python -m venv .venv
    source .venv/Scripts/activate
    pip install -r requirements.txt

1. 1.Download Data: Download the MAESTRO Dataset v3.0.0 and place the maestro-v3.0.0 folder inside the data/raw/ directory.

Try It

A standalone web app (custom HTML/CSS/JS frontend in web/, served by a small FastAPI backend in server/main.py) wraps the trained DisentangledVAE so you can recombine rhythm/pitch between MIDI clips without writing any code. See [QUICKSTART.md](QUICKSTART.md) for how to launch it on Windows with one double-click.

<!--The original Streamlit version (app.py) still works as a fallback - both share the same model pipeline in src/inference/, so they can't drift apart in behavior.

A DAW plugin version of the same tool exists as unbuilt source in [plugin/](plugin/README.md) - see plugin/BUILD.md to compile it yourself.--!>