One Hot Chord

An experimental deep learning project for real-time chord recognition from audio input. This project demonstrates how machine learning can be applied to music analysis. You can access the demo at https://onehotchord.com

Note: This is an experimental project and may not work perfectly in all scenarios.

Features

• Real-time chord recognition from audio input
• Browser-based interface with visual feedback
• Privacy-focused design - all processing happens on your device
• Supports common chord types: major, minor, diminished, 7th chords
• Responsive visualization of detected notes

How It Works

One Hot Chord uses a deep neural network to analyze audio and identify chords:

1. Captures audio from your microphone
2. Extracts frequency information using a Constant-Q Transform
3. Processes these features through a neural network
4. Identifies the root note and chord type
5. Displays the results in real-time

Project Structure

• gen_samples.py - Generates synthetic chord samples for training
• preprocess.py - Extracts features from audio samples
• train.py - Trains the neural network model
• model.py - Defines the neural network architecture
• docs/ - Web demo

Quick Start

Prerequisites

• Python 3.8+
• FluidSynth and a SoundFont file (for sample generation)

Setup

1. Clone the repository
2. Install dependencies: pip install -r requirements.txt
3. Download a SoundFont file (e.g., FluidR3_GM.sf2) to the sf2/ directory

Training Pipeline

# Generate training samples
python gen_samples.py

# Preprocess audio samples
python preprocess.py [wav files]

# Train the model
python train.py chord_dataset.npz

# Try real-time recognition
python listen.py

Web Interface

Serve the web interface:

cd www
python -m http.server 8000

Then open your browser to http://localhost:8000

Limitations

This is still an experimental project built in a weekend. It has limitations:

• Works best with clean audio input
• May struggle with complex chord voicings
• Limited to the chord types it was trained on
• Performance varies depending on audio quality and background noise

License

This project is licensed under the Apache License - see the LICENSE file for details.

Acknowledgments

• FluidSynth for MIDI synthesis
• Librosa for audio processing
• PyTorch for neural network implementation
• ONNX Runtime for model deployment