The Strummin' Dataset: an international pop/rock curated audio selection for strumming patterns recognition

Description The Strummin’ Dataset is a curated collection of annotated pop/rock music excerpts developed as part of the Strummin’ Gym project, which aims to support beginner guitarists in acquiring core rhythmic accompaniment skills. Unlike existing datasets for beat, tempo, or onset detection, this dataset addresses the novel task of categorical strumming pattern recognition, focusing on mid-level rhythmic structures (binary, quaternary, ternary) widely used in pop/rock pedagogy. Contents and Structure The release consists of: A dataset/ folder containing: Original .csv file with metadata and annotations Downloaded audio files (YouTube audio .wav files, 2chs., 48kHz) Pre-computed rhythmic features (w. LibROSA, Essentia and RP_extract) in .npz format, .svg high-resolution plots for each feature/file Supporting scripts: data_processing.py: feature_extractor class implemented and features extraction pipeline data.py: dataset and dataloader classes definition with PyTorch and LightningDataModule integration Annotations Each song entry in the .csv file includes: Title and Performer YouTube URL of the song Tempo Annotations: bpm_marco: manually validated bpm_tunebat: from Tunebat (Essentia.js) bpm_songBPM: from SongBPM (Spotify API) Strumming Pattern Label: 0: binary articulation 1: quaternary articulation 2: ternary articulation Download Flag: indicating successful pre-retrieval Statistics title performer yt_url bpm_marco bpm_tunebat bpm_songbpm label downloaded Through her eyes Dream Theater https://www.youtube.com/watch?v=3gy8hkVMqwQ 57 114 114 0 True Rehab Amy Winehouse https://www.youtube.com/watch?v=KUmZp8pR1uc 145 72 72 1 True Let's twist again Chubby Checker https://www.youtube.com/watch?v=Frw7U5X8dhg 167 168 168 2 True Total samples: 257Average duration (per file): 242.78 seconds (around 4min. per song)Total contents duration: 17:19:53 (hh:mm:ss) Samples (per class): Class 0: 89 samples Class 1: 99 samples Class 2: 69 samples Usage The dataset can be used for training and evaluating deep learning models in: Strumming pattern classification Rhythmic feature analysis & recognition (BPM, onsets retrieval, rhythmic segmentation etc.) Educational/MIR applications It integrates with PyTorch and TorchAudio, supporting waveform-based and descriptor-based pipelines. Segment-wise data loading and label inheritance allow flexible experiments with pre-defined (but customizable) 10-second clips (inspired by AudioSet "standard"). References "M. Pennese, S. Giacomelli and C. Rinaldi A Kolmogorov Arnold Network NAS Framework for Strumming Pattern Recognition in Technology-Enhanced Pop/Rock Music Education" (under peer-review for IS2 IEEE International Symposium of The Internet of Sounds 2025)