Musical Key CNN

This repository provides a full pipeline for musical key classification based on convolutional neural networks, inspired by the paper [1]. It contains scripts for preprocessing, training, evaluation, and prediction of key labels on new music tracks, using the GiantSteps and GiantSteps-MTG datasets and the Camelot Wheel for key mapping.

---

Table of Contents

• Description
• Setup and Installation
• Key Prediction for Your Own Songs
• Dataset Preparation
• Preprocessing
• Training
• Evaluation
• Literature

---

Description

This repository implements a CNN model for musical key detection. It provides scripts to:

• Preprocess datasets (extract CQT spectrograms, prepare annotations, augment with pitch shifts)
• Train the model from scratch
• Evaluate model performance with MIREX key evaluation metrics
• Predict keys for custom .mp3 files

---

Setup and Installation

We recommend using a Python virtual environment for reproducibility and package management:

python3 -m venv venv
source venv/bin/activate

Install the required packages:

pip install -r requirements.txt

Important

For PyTorch and torchaudio, follow the instructions at https://pytorch.org/get-started/locally/. Choose the correct command for your CUDA/cuDNN or CPU environment.

Key Prediction for Your Own Songs

You can analyze any individual .mp3 or a folder of .mp3 tracks using the provided model or your own trained model:

python predict_keys.py -f path/to/your_song.mp3
python predict_keys.py -f path/to/your/music_folder/

The script prints a summary table with:

• Filename
• Classification index (0-23)
• Index according to the Camelot Wheel (e.g., "8A" or "3B")
• The corresponding key

You can set the model checkpoint path with -m path/to/your_model.pt and the computation device with --device cuda or --device cpu.

Dataset Preparation

For training and evaluation, you need the following datasets:

• GiantSteps MTG Key Dataset (Training)
• GiantSteps Key Dataset (Evaluation)

Directory structure: Place or symlink the datasets under the Dataset/ folder:

Dataset/
    giantsteps-key-dataset/
    giantsteps-mtg-key-dataset/

Preprocessing

Before training or evaluation, preprocess the datasets to generate CQT spectrograms for all tracks and pitch-shifted variants.

python preprocess_data.py

All resulting .pkl spectrograms are stored in subfolders of Dataset/.

Training

To train a new key classification model on the MTG dataset, run:

python train.py

You can modify hyperparameters or training parameters by editing train.py.

Evaluation

To evaluate a trained model (e.g., calculate MIREX scores on GiantSteps):

python eval.py

The output includes overall accuracy and weighted MIREX scores.

The following table contains the percentage ratios and the weighted Mirex scores:

Method	Weighted	Correct	Fifth	Relative	Parallel	Other
keynet.pt	73.51	66.72	8.11	6.79	3.48	14.90
Mixed In Key 8.3	75.70	69.37	8.11	5.13	3.64	13.74
RekordBox 7.12	65.53	56.79	11.92	5.96	4.97	20.36

Literature

Please cite and refer to the original publication for scientific use and further reading:

• [1] F. Korzeniowski and G. Widmer. "Genre-Agnostic Key Classification With Convolutional Neural Networks". In: Proceedings of the 19th International Society for Music Information Retrieval Conference (ISMIR) (2018) arXiv

• [2] F. Korzeniowski and G. Widmer. "End-to-End Musical Key Estimation Using a Convolutional Neural Network". In: Proceedings of the 25th European Signal Processing Conference (2017) arXiv