Euterpe

Euterpe is a generative system that composes music based on artificial intelligence models. It is part of the Master's Thesis "Generación de Música Personalizada a través de Modelos Generativos Adversariales" by Rafael Luque Tejada (2025).

The system allows training and generation of music using Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN), based on labeled audio datasets.

Features

• Music generation conditioned by genre.
• Training using VAE or GAN architecture.
• Dataset filtering through a list of valid files.
• Input normalization and reconstruction through inverse STFT.
• Genre-aware decibel scaling.
• Modular design prepared for experimentation.

Requirements

• Python 3.12.7
• librosa, torch, pandas, numpy, soundfile, matplotlib, and others (see requirements.txt)
• CUDA-compatible GPU (optional but recommended)

Installation

Clone this repository and install the dependencies:

git clone https://github.com/yourusername/euterpe.git
cd euterpe
pip install -r requirements.txt

Usage

The entry point is main.py. The script expects three arguments:

python main.py <datasets_path> <valid_files_csv_path> <model_to_train>

Arguments

• datasets_path: path to the directory containing the dataset (audio files).
• valid_files_csv_path: path to the CSV file listing valid files for training.
• model_to_train: vae or gan, depending on the model to train.

Example

python main.py ./data ./valid_files.csv vae

Project Structure

euterpe/
├── main.py                # Entry point
├── models/                # Model architectures: VAE, GAN, etc.
├── datasets/              # Dataset utilities and preprocessing
├── training/              # Training loops and evaluation
├── generation/            # Audio generation logic
├── config/                # Environment and parameter configuration
└── utils/                 # Helpers (logging, audio tools...)

License

This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.

---

© 2025 Rafael Luque Tejada — lukemaster.master@gmail.com