AI Live Drum Kit is a Python-based project designed to add realistic, dynamic drum accompaniments to live MIDI input. This tool enables real-time drumming experiences driven by AI, without requiring a GPU. It also includes utilities to train custom models if desired.

• Live Drum Accompaniment: Generates and plays drum patterns in real-time, based on live MIDI input.
• Model Training Utilities: Train custom drum accompaniment models in 20-60 epochs with a simple configuration.
• Efficient and Lightweight: No GPU needed! The model operates smoothly on a standard CPU.
• User-Friendly MIDI Integration: Uses mido for MIDI handling and Keras for AI model building.

• Python 3.8+
• Dependencies (install via pip):

  pip install numpy mido keras

• Configuration: Ensure all required variables are correctly set up in the config.py file. (MIDIS is mandatory only for training, the rest have defaults.)

Clone the repository and navigate to the directory:

git clone https://github.com/squidKid-deluxe/ai-live-drumkit.git
cd ai-live-drumkit

Install the required packages:

pip install numpy mido keras

The AI Live Drum Kit project uses a neural network model built in Keras to generate drum accompaniments. The model is trained on MIDI files to learn the nuances of drumming patterns.

1. Specify where your MIDI training files are using the MIDIS variable in config.py.

2. Run the training command:

   python main.py train

3. The model will save after training, use load as a command line argument to load the saved model for retraining or use.

To start the live AI accompaniment:

python main.py live load

Your MIDI inputs will now be enhanced with AI-generated drum patterns in real-time!

For a quick demonstration, add MIDI files and train the model with 20 epochs. Then, start the live mode to hear drum accompaniment on MIDI input.

• CPU Timing: 20 epochs on 40 MIDI files takes about 5-10 minutes on a 3.4 GHz quad-core CPU.
• GPU Timing: Unknown but expected to be faster.

• Add support for more complex drum patterns.
• Optimize the model for faster inference.
• Less overfitting (but is that actually bad in this case?)

This project is licensed under the MIT License. See the LICENSE file for details.