The CP-WTC Dataset

The CP-WTC Dataset provides valence-arousal annotations for excerpts from recordings of The Well-Tempered Clavier (WTC), Book I by J. S. Bach, performed by six famous pianists. Book I of the WTC comprises 48 pieces in total, 24 preludes and 24 fugues covering all major and minor keys. Following the annotation strategy of Battcock & Schutz [1], and to ensure compatibility of this dataset with their data, only the first eight bars of each recording were annotated. Each audio excerpt was annotated by several human listeners, resulting in 22 to 29 annotations per excerpt. Robust mean values for valence and arousal were then calculated by removing outliers using the Minimum Covariance Determinant (MCD) estimator. Overall, this results in 288 annotated audio segments (48 pieces x 6 performances), together with 7,652 individual listener annotations. Due to copyright restrictions, the original audios, taken from commercial CD recordings, are not included. Instead, the dataset contains various extracted features corresponding to each recording. Information about the original recordings is documented for reference. The following is the structure of this dataset: Metadata: Recording Details Description for Annotation Process Feature Details Metadata & Valence-Arousal Annotation Features: Audio Low-Level / Mid-Level Features BPM Symbolic: Scores Alignments MIDI The latest version of this dataset and any future updates are also maintained on GitHub. This dataset was first introduced in the following paper: @inproceedings{Chowdhury2021OnPE, title={On Perceived Emotion in Expressive Piano Performance: Further Experimental Evidence for the Relevance of Mid-level Perceptual Features}, author={Shreyan Chowdhury and Gerhard Widmer}, booktitle={Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)}, year={2021} } If you use this dataset in your research, please cite the following paper: @inproceedings{ching2026PerfRLT, title={Learning to Predict Performance-Induced Emotion Differences in Classical Piano Music}, author={Joann Ching and Gerhard Widmer}, booktitle={Proceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR)}, year={2026}, address={Abu Dhabi, UAE} } Recordings Recordings used in this dataset are directly acquired from official recordings of the pianists. The "bibliographic" data can be found in ./metadata/recording_details.txt. Recordings from 6 pianists are included; in parentheses are their abbreviations (pianist_id) used throughout the dataset: Glenn Gould (GG) Friedrich Gulda (FG) Angela Hewitt (AH) Sviatoslav Richter (SR) András Schiff (AS) Rosalyn Tureck (RT) Annotations The participants in the annotation are CS/AI students of a course at the Johannes Kepler University, without specific musical background. Each participant heard a subset of the recordings (all 48 pieces played by one pianist), and could listen to a recording as many times as they like. A documentation of the annotation process can be found in ./metadata/annotation_descriptions.pdf, which includes (1) the definition of valence/arousal given to the annotators, and (2) the annotation setup and instructions presented to them. Only the first 8 bars of the recordings are used for collecting the valence-arousal annotations. If you have the same recordings used in this dataset, the timestamps for creating the annotation segments can be found in ./metadata/excerpt_timestamps.txt. Run ./utils/recreate.py to extract the segments for annotation. You can find the valence-arousal annotations of each recording in ./metadata/metadata_annotations.csv, along with the raw annotations from the annotators. The raw annotations range of valence and arousal are as follows: Arousal: 0~100 (in steps of 10, a total of 11 levels) Valence: -5~+5 (in steps of 1, a total of 11 levels) Audio Features Instead of the original audio files of the recordings, we provide several sets of features extracted with various audio and music analysis packages. Feature metadata for all snippets can be found at ./metadata/metadata_audio_features.csv. Respective scripts for feature extraction can be found under ./utils/. ./features/lowlevel: contains features computed by Essentia and Librosa from audio waveform or spectrum. The features are extracted in the unit of "snippets", which are 10-second chunks from the original recording (until the last full 10 seconds). $./features/lowlevel ├── Gould ├── {piece_num}-{snippet_num}-essentia_dissonance.pkl ├── {piece_num}-{snippet_num}-essentia_dynamic_complexity.pkl ├── ... ├── {piece_num}-{snippet_num}-librosa_bpm.pkl └── ... ├── Gulda └── ... ├── ... └── Tureck └── ... ./features/midlevel: refer to this paper for the definitions of each of these features. The Midlevel Features Dataset was used to train the model with domain adaptation for piano music, and the trained model was used to predict these features for this dataset [2]. $./features/midlevel ├── Gould ├──{piece_num}-{snippet_num}-midlevel_features.pkl └──... ├── Gulda └── ... ├── ... └── Tureck └── ... ./features/bpm: pickle files of the BPM extracted by librosa. $./features/bpm ├── Gould ├──{piece_num}-{snippet_num}-librosa_bpm.pkl └──... ├── Gulda └── ... ├── ... └── Tureck └── ... Score Features are computed with Partitura directly from the musical score (./symbolic/scores/**.musicxml) of the pieces instead of the audio files. The score information and the audio files were linked using automatic score-to-performance alignment. ./features/snippets_melspec contains the corresponding mel-spectrograms for each snippet. See ./utils/generate_melspecs.py for the associated script. $./features/snippets_melspec ├── Gould ├── {piece_num}-{snippet_num}.png ├── {piece_num}-{snippet_num}.png ├── ... ├── {piece_num}-{snippet_num}.png └── ... ├── Gulda └── ... ├── ... └── Tureck └── ... Symbolic Resources Scores The scores for each of the 48 pieces from The Well-Tempered Clavier can be found under ./symbolic/scores. The scores are manually reviewed and corrected to match the Henle Verlag Urtext edition (./symbolic/henle). Note that: each of the scores is universal across the performers; therefore, the files are named with only the piece_num. $./symbolic/scores ├── 01.musicxml ├── 02.musicxml ├── ... └── 48.musicxml Alignments Alignments from each of the performance audios to the scores are also provided under ./symbolic/alignments. $./symbolic/alignments ├── Gould ├──{piece_num}.txt └──... ├── Gulda └── ... ├── ... └── Tureck └── ... MIDI In line with the related data of the audio recordings, the performance MIDI data are also included in this dataset. Note: These files are only transcription results from current SOTA models, instead of the ground truth directly recorded from pianists. MIDI files are transcribed by the Transkun model. References [1] Aimee Battcock & Michael Schutz. "Acoustically Expressing Affect." Music Perception: An Interdisciplinary Journal. 37. pp. 66-91, 2019. [journal article] [2] Shreyan Chowdhury and Gerhard Widmer. "On perceived emotion in expressive piano performance: Further experimental evidence for the relevance of mid-level perceptual features." Proceedings of the 22nd International Society for Music Information Retrieval Conference (ISMIR), 2021. [paper] [3] Joann Ching and Gerhard Widmer. "Learning to Predict Performance-Induced Emotion Differences in Classical Piano Music." Proceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR), 2026. [paper] [4] Shreyan Chowdhury and Gerhard Widmer. "Decoding and Visualising Intended Emotion in an Expressive Piano Performance." in Extended Abstracts for the Late-Breaking Demo Session of the 23rd International Society for Music Information Retrieval Conference (ISMIR), 2023. [paper] [5] Shreyan Chowdhury. "Modelling Emotional Expression in Music Using Interpretable and Transferable Perceptual Features." Ph.D. Thesis, Johannes Kepler University Linz, Linz, Austria, 2022. [paper] License The CP-WTC Dataset is released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). It is provided primarily for research purposes and is prohibited from being used for commercial purposes. Acknowledgement This work is supported by the European Research Council (ERC) under the EU’s Horizon 2020 research & innovation programme, grant agreement No. 670035 (“Con Espressione”) and No. 101019375 (“Whither Music?”), and the Federal State of Upper Austria (LIT AI Lab).