Multi-Keyboard 36-Class Piano Chord Dataset

This dataset contains raw audio samples of piano chords designed for training and evaluating automatic chord recognition systems. The dataset consists of 36 distinct chord classes, representing the combination of 12 chromatic root notes (C, C#, D, D#, E, F, F#, G, G#, A, A#, B) and 3 chord qualities (major, minor, and diminished). All chords are played within the fourth octave. The inclusion of diminished chords provides a benchmark for testing classifier sensitivity on harmonically dense and less frequently occurring chord types. The audio files are completely raw, without any post-processing filters, to preserve authentic spectral characteristics. Dataset Generator Tool: sglkc/chords-dataset-generator Technical Specifications Format: OGG/OPUS Sample Rate: 48,000 Hz Bit Depth: 32-bit (Mono) Total Samples: 8,640 files Dataset Structure 1. Training Dataset (7,200 samples) Distribution: 200 samples per class. Variability: Randomized note attack delays of 0–150 ms to simulate natural human playing and prevent overfitting. Keyboard: Techno T-9890i keyboard with Acoustic Grand Piano voice. Recording Hardware: Xiaomi Pad 2. 2. Evaluation Dataset (1,440 samples) Distribution: 40 samples per class. Variability: Randomized note attack delays of 0–100 ms. Half of the samples per class are recorded at a higher volume to test model robustness against dynamic changes. Keyboard: Yamaha PSR-S910 keyboard with Grand Piano voice. Recording Hardware: ROG Flow X13. Class Labels Reference The 36 classes follow a standard naming convention combining the root note, quality, and octave (e.g., C_major_4, C_minor_4, C_diminished_4, C#_major_4). Audio samples are grouped by the respective class directories. Archive Structure training├── C_major_4│ ├── C_major_4-1.ogg│ ├── ...│ ├── C_major_4-200.ogg├── C#_diminished_4│ ├── C#_diminished_4-1.ogg│ ├── ...│ ├── C#_diminished_4-200.ogg 36 directories, 7200 files Citation If you use this dataset or the generator tool in your research, please properly cite it with the Zenodo record (10.5281/zenodo.21077144).