---
license: cc-by-nc-sa-4.0
tags:
- music
- piano
- midi
- expressive performance
- score-to-performance
pretty_name: Piano Expression Refined Score and Performance MIDI Dataset
size_categories:
- 10K<n<100K
---


# PERiScoPe: Piano Expression Refined Score and Performance MIDI Dataset

**PERiScoPe** is a large-scale dataset of aligned piano scores and performances collected and utilized in the paper [**SyMuPe: Affective and Controllable Symbolic Music Performance**](https://arxiv.org/abs/2511.03425).

The dataset combines and processes open-source collections such as [(n)ASAP](https://github.com/CPJKU/asap-dataset) and [ATEPP](https://github.com/tangjjbetsy/ATEPP) with a set of curated web-collected and transcribed MIDI performances, all aligned and cleaned to produce parallel score and performance note sequences.

## Related Resources

*   **GitHub:** https://github.com/ilya16/SyMuPe
*   **Website:** https://ilya16.github.io/SyMuPe

## Dataset Structure

The repository is organized by version. 

```text
/
├── metadata.csv                # Metadata for v1.1 (copy of v1.1/metadata.csv)
├── DISCLAIMER.md               # Usage terms and copyright notice
│
├── v1.1/                       # CURRENT (Bug-fix Release)
│   ├── periscope_raw_v1.1.tar.gz
│   ├── periscope_aligned_v1.1.tar.gz
│   └── metadata_v1.1.csv
│
└── v1.0/                       # LEGACY (SyMuPe Paper Snapshot)
    ├── periscope_raw_v1.0.tar.gz
    ├── periscope_aligned_v1.0.tar.gz
    └── metadata_v1.0.csv
```

### Archives Content
*   **`_raw_`**: Contains the input data grouped by composer, piece name and (if available) movement name.
    *   **Score:** MusicXML files and MIDI scores computed using [**partitura**](https://github.com/CPJKU/partitura).
    *   **Performance:** MIDI files taken from ASAP, ATEPP or transcribed from audio using [**Transkun V2**](https://github.com/Yujia-Yan/Transkun).
*   **`_aligned_`**: Contains the processed pairs.
    *   **Score:** MIDI scores with tracks merged into a single track.
    *   **Performance:** MIDI files with note-by-note alignment to score MIDI files. The alignment was computed using [**Parangonar**](https://github.com/sildater/parangonar), cleaned to remove high inter-onset/intra-onset deviations, and interpolated for all unperformed notes.
    *   **Alignment:** numpy array file with the alignment between the score and performance MIDI files.

### Unarchived File Structure

After extracting the archives, the dataset follows this directory structure:

```text
PERiScoPe-1.1/
├── metadata.csv
├── data/
│   ├── Johann_Sebastian_Bach/
│   │   ├── Italian_Concerto,_BWV_971/
│   │   │   ├── ASAP_KyykhynenT03.mid
│   │   │   ├── ...
│   │   │   ├── score_ASAP.mid
│   │   │   └── score.musicxml
│   │   └── ...
│   └── ...
└── data_aligned/
    ├── Johann_Sebastian_Bach/
    │   ├── Italian_Concerto,_BWV_971/
    │   │   ├── ASAP_KyykhynenT03.mid
    │   │   ├── ASAP_KyykhynenT03.npy
    │   │   ├── ...
    │   │   └── score_ASAP.mid
    │   └── ...
    └── ...
```

Using the prefixes `ASAP_`, `ATEPP_`, and `PERiScoPe_` in the filenames, it is possible to filter out MIDI performances from specific source collections.

## Version History

### **v1.1 (January 2026)**
A maintenance release improving data quality.
*   **Score Fix:** Recomputed score MIDI files to match MusicXML note velocities (fixing the previous `velocity=1` issue).
*   **Metadata Fix:** Corrected metadata entries for 318 items.
*   **Alignment Update:** Applied an updated alignment cleaning algorithm with more rigorous filtering of low alignments. This results in 1,162 fewer pairs than v1.0 but ensures higher alignment accuracy.

#### Statistics

| Dataset | Composers | Scores | Performances | Duration | Notes |
| :--- | :--- | :--- | :--- | :--- | :--- |
| (n)ASAP | 16 (16) | 215 (221) | 992 (1,066) | 87h (94h) | 3.3M (3.6M) |
| ATEPP | 19 (25) | 652 (1,553) | 7,243 (11,674) | 603h (1,003h) | 18.5M (31.9M) |
| Web-Collected | 66 (82) | 1,151 (2,301) | 27,580 (28,500) | 2,158h (2,686h) | 66.5M (82.0M) |
| **Total** | **66 (82)** | **1,163 (2,738)** | **35,815 (46,473)** | **2,848h (3,784h)** | **88.4M (117.5M)** |

*: Format: **Aligned (Raw)**. "Aligned" refers to the note-aligned pairs in `_aligned_`. "Raw" refers to the total source material in `_raw_`.

### **v1.0 (March 2025) - Legacy**
*   The original dataset used for the experiments and training the models presented in the SyMuPe paper.
*   Retained for reproducibility of the results reported in the paper.
*   **Known Issues:** Score MIDI files might have note velocities of 1. Some performances correspond to incorrect compositions.

#### Statistics

| Dataset | Composers | Scores | Performances | Duration | Notes |
| :--- | :--- | :--- | :--- | :--- | :--- |
| (n)ASAP | 16 (16) | 215 (221) | 1,004 (1,066) | 89h (94h) | 3.4M (3.6M) |
| ATEPP | 19 (25) | 656 (1,553) | 7,794 (11,674) | 664h (1,003h) | 21.0M (31.9M) |
| Web-Collected | 67 (82) | 1,162 (2,301) | 28,179 (33,733) | 2,215h (2,686h) | 68.7M (82.0M) |
| **Total** | **67 (82)** | **1,176 (2,738)** | **36,977 (46,473)** | **2,968h (3,784h)** | **93.1M (117.5M)** |

## Copyright and Licensing

**Usage License:** The dataset structure, metadata, and alignment annotations are distributed under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/deed.en) license.

**Disclaimer:** The MIDI files contained in this dataset are derived from various sources, including public web scraping and automatic transcription. We do not claim ownership of the underlying musical compositions or the original audio performances. These files are provided solely for non-commercial research purposes.

By accessing this dataset, you agree to the terms outlined in [**DISCLAIMER.md**](https://huggingface.co/datasets/ilya16/PERiScoPe/blob/main/DISCLAIMER.md).

## Citation

If you use the dataset, please cite the paper:

```bibtex
@inproceedings{borovik2025symupe,
  title = {{SyMuPe: Affective and Controllable Symbolic Music Performance}},
  author = {Borovik, Ilya and Gavrilev, Dmitrii and Viro, Vladimir},
  year = {2025},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
  pages = {10699--10708},
  doi = {10.1145/3746027.3755871}
}
```