---
title: 'GigaMIDI: Expressive MIDI Performance Dataset'
url: https://www.emergentmind.com/topics/gigamidi
type: topic
---

# GigaMIDI: Expressive MIDI Performance Dataset

GigaMIDI is a large-scale symbolic music dataset introduced as both a consolidated MIDI corpus and a framework for expressive music performance detection. It was presented in "The GigaMIDI Dataset with Features for Expressive Music Performance Detection" [2502.17726] as containing **1,437,304 unique MIDI files**, **5,334,388 MIDI instrument tracks**, and **1,824,536,824 MIDI note events**, with the name referring to the fact that it contains more than \(10^9\) note events. The dataset is positioned for music information retrieval (MIR), computational musicology, and generative music research, while its most distinctive methodological contribution is a set of track-level heuristics for distinguishing **non-expressive** MIDI tracks from **expressive** ones.

## 1. Corpus scale and internal composition

GigaMIDI is described as the **largest MIDI collection available for research purposes under fair dealing** at the time of publication [2502.17726]. The paper reports that the collection contains **more than 40,000 hours** of music and compares this scale with **436,631 files** for MetaMIDI and **174,533 files** for Lakh MIDI. The initial collected version had **1,773,996 MIDI files**, and about **20%** of the dataset underwent cleaning, including deduplication using **MD5 checksums**.

| Attribute | Value |
|---|---|
| Unique MIDI files | 1,437,304 |
| MIDI instrument tracks | 5,334,388 |
| MIDI note events | 1,824,536,824 |
| Single-track files | 56.8% |
| Multi-track files | 43.2% |
| Drum tracks | 996,164 |
| Non-drum tracks | 4,338,224 |

The dataset is divided into three primary subsets: **“all-instrument-with-drums” (22.78%)**, **“drums-only” (56.85%)**, and **“no-drums” (20.37%)**. The drums-only files are described as typically short and loop-like, often under 50 bars, whereas piano-based files span a broader duration range, often 10 to 300 bars. This suggests that GigaMIDI is not a uniform corpus of complete multitrack compositions, but an aggregate symbolic collection with a substantial proportion of loop-oriented material.

Instrument-group note-event statistics are strongly concentrated. **Piano accounts for 60.2% of note events** and **drums for 17.4%**, together making up **77.6%** of all note events. The remaining groups are smaller: chromatic percussion **2.4%**, organ **1.8%**, guitar **6.7%**, bass **4.2%**, strings **1.1%**, ensemble **2.1%**, brass **0.7%**, reed/pipe **1.1%**, synth lead **0.5%**, synth pad **0.6%**, synth FX **0.3%**, ethnic **0.3%**, percussive FX **0.3%**, and sound FX **0.249%**. The authors relate the piano dominance to MIDI practice itself: keyboards are common input devices, piano tracks are often used as compositional sketches, and piano writing tends to be dense and polyphonic.

Additional descriptive statistics reinforce the corpus’s breadth. The number of notes per file between the **5th and 95th percentiles** ranges from **13 to 931 notes**. Common **ticks per quarter note (TPQN)** values between the **5th and 95th percentiles** range from **96 to 960**, with **480** and **960** especially common. The dataset also includes **7,108,181 loops** of lengths **1 to 8 bars**, extracted from tracks using an adapted loop detection algorithm.

## 2. Acquisition, cleaning, standardization, and metadata

The collection process was aggregative and partly manual [2502.17726]. The authors gathered freely available MIDI files from sources such as **Zenodo**, **GitHub**, and other public MIDI repositories via web scraping. They state that **over 50%** of GigaMIDI was acquired by their own web scraping and organization efforts, while the remainder incorporated previously available public MIDI datasets. Files were organized and deduplicated by comparing **MD5 hash values**.

A central preprocessing step was standardization to the **General MIDI (GM)** specification. The paper states that non-GM drum tracks were remapped to GM, and that single-track drum files incorrectly allocated to MIDI channel 1 were re-encoded to channel 10 to avoid being misidentified as piano tracks. All drum tracks were further standardized through GM drum remapping. The dataset is hosted on the **Hugging Face Hub**, with raw metadata there as well, and the paper notes this as useful for symbolic music workflows such as parsing and tokenization with **Symusic** and **MidiTok**.

Metadata enrichment is present but partial. The dataset includes style metadata aligned to the Musicmap topology. In total, **195,737 files** have style annotations, with prevalent styles including classical, pop, rock, and folk. To improve metadata quality, a subset of **29,713 MIDI files** was manually reviewed and style-annotated by a doctoral-level music researcher, and **9,980** audio-text-matched style metadata entries from MetaMIDI were manually remapped.

The access model is explicitly legalistic. GigaMIDI is distributed under **Fair Dealing provisions for research and private study under the Canadian Copyright Act**. On Hugging Face, access is restricted to research purposes only, and the platform collects the user’s legal name and email as proof of agreement. Filenames are anonymized using **MD5 hash encryption**. This legal framing is part of the dataset’s identity: it is presented as research-oriented, accessible, and interoperable, but not as an unrestricted public-domain redistribution.

## 3. Expressive performance detection as a track-level problem

A defining feature of GigaMIDI is that it treats expressive-performance detection as a **track-level** problem rather than only a file-level one [2502.17726]. The paper argues that MIDI files usually do not encode metadata identifying whether tracks are expressive or non-expressive, even though this distinction matters for MIR, computational musicology, and expressive symbolic generation.

The authors define **non-expressive** tracks as those with relatively fixed velocity levels and quantized rhythms, yielding metronomic and mechanical results. **Expressive** tracks, by contrast, exhibit **microtiming variation** relative to an ideal metric grid and **dynamic variation** reflected in changing MIDI velocities. They organize the resulting space into four conceptual classes:

- **NE**: non-expressive  
- **EO**: expressive in onset only  
- **EV**: expressive in velocity only  
- **EP**: expressive in both onset and velocity  

The first heuristic family extends simple distinct-value counts. Let  
\[
\mathbf{x} = [x_1, \dots, x_n]
\]
be either the list of note velocities in a track or the list of note onset deviations from a quantized grid. The number of unique values is counted as \(c_{velocity}\) or \(c_{onset}\), and then normalized as percentages:
\[
\mathbf{c_{velocity-ratio} = c_{velocity} \div 127 \times 100}
\]
\[
\mathbf{c_{onset-ratio} = c_{onset} \div TPQN \times 100}
\]

In the paper’s terminology, \(c_{velocity-ratio}\) is the basis of **DNVR** and \(c_{onset-ratio}\) is the basis of **DNODR**. The stated interpretation is straightforward: expressive tracks should use a wider set of distinct velocity values and a larger variety of onset deviations than mechanically entered or tightly quantized tracks. The normalization by **127** or **TPQN** is intended to reduce sensitivity to track length and to improve comparability across files with different timing resolutions.

On the full corpus, the baseline distinct-value thresholds yielded **NE: 62.5%**, **EO: 7.2%**, **EV: 27.4%**, and **EP: 2.9%**. Using **DNVR** and **DNODR** thresholds instead, the distribution became **NE: 52.3%**, **EO: 9.1%**, **EV: 24.2%**, and **EP: 14.4%**. This indicates that normalization materially changes the inferred prevalence of expressive onset behavior.

## 4. Note Onset Median Metric Level (NOMML)

The second major heuristic, **Note Onset Median Metric Level (NOMML)**, is the paper’s best-performing method [2502.17726]. Rather than counting the number of distinct onset deviations, NOMML asks how fine-grained note onsets are relative to a hierarchy of duple and triplet metric grids. If most onsets fall on coarse positions such as quarter-note or eighth-note locations, the track is likely to be quantized; if the median onset lies only on much finer subdivisions, or outside all considered regular grids, the track is treated as more expressive.

The paper defines a metric hierarchy using both **duple** and **triplet** onset grids, with parameter \(k\) controlling the maximum depth. For the reported experiments, **\(k = 6\)**, which yields duple and triplet levels up to **128th-note granularity**. For each onset \(o_i\), the algorithm first tries duple periodicities
\[
p \gets \frac{TPQN}{2^j}
\]
for \(j = 0, \dots, k-1\). If
\[
o_i \bmod p \equiv 0
\]
the onset is assigned level
\[
2j
\]
and the search stops. If no duple level matches, the algorithm tries triplet periodicities
\[
p \gets \frac{2 \cdot TPQN}{3 \cdot 2^j}
\]
and if
\[
o_i \bmod p \equiv 0
\]
the onset is assigned level
\[
2j + 1.
\]
If neither duple nor triplet levels match up to depth \(k-1\), the onset is assigned the extra category
\[
2k.
\]
The track-level heuristic then returns
\[
\text{median}(\mathbf{c}).
\]

The notation is interleaved so that duple levels are encoded as even numbers,
\[
dup^i = 2i,
\]
and triplet levels as odd numbers,
\[
trip^i = 2i + 1.
\]
With \(k=6\), possible values range from 0 up to 12, where **12** is the “beyond grid” category. The paper’s decisive claim is that high NOMML values, especially level 12, indicate that the median onset lies outside all considered regular duple/triplet positions, suggesting expressive microtiming.

The paper provides an intuitive example with \(TPQN = 120\). An onset at tick 60 aligns with both duple and triplet grids, but because duple is prioritized, its metric level is \(dup^1\). An onset at tick 40 aligns only with triplet levels, so its metric level is \(trip^1\). This design makes the median metric level an interpretable summary of how strongly a track departs from regular quantization.

## 5. Evaluation protocol and construction of the expressive subset

Evaluation was carried out on a manually curated binary ground-truth set of **361 NE tracks** and **361 EP tracks**, for a total of **722 instances** [2502.17726]. The non-expressive examples came from **ASAP score tracks**. The expressive examples came from **ASAP performance tracks**, the **Vienna 4x22 Piano Corpus**, **Saarland Music Data**, **Groove MIDI**, and the **Batik-plays-Mozart Corpus**. Only tracks for which the authors had **100% confidence** in the label were included. The classes were intentionally balanced; otherwise expressive tracks would have outnumbered non-expressive ones by roughly **10 to 1**.

The curated set was split into **80% training** for threshold selection and **20% testing** for final evaluation. Threshold selection used **logistic regression** together with **leave-one-out cross-validation (LOOCV)** on the training portion. The paper uses the **P4** metric:
\[
P_4 = \frac{4 \cdot TP \cdot TN}{4 \cdot TP \cdot TN + (TP + TN) \cdot (FP + FN)}
\]

| Heuristic | Optimal threshold | Test accuracy |
|---|---:|---:|
| Distinct Velocity | 52 | 77.9% |
| Distinct Onset | 42 | 77.9% |
| DNVR | 40.965% | 83.4% |
| DNODR | 4.175% | 98.2% |
| NOMML | level 12 | 100% |

The corresponding P4 values for threshold selection were **0.7727** for Distinct Velocity, **0.7225** for Distinct Onset, **0.7727** for DNVR, **0.9529** for DNODR, and **0.9952** for NOMML. The paper also reports confusion-related quantities and notes that onset-based heuristics generally outperformed velocity-based ones. One reason given is that even non-expressive score tracks may encode traditional dynamics using varying MIDI velocities, whereas onset quantization remains much more discriminative.

After evaluation, the authors curated an expressive version of GigaMIDI using **NOMML level 12** as the threshold. In the full dataset, the NOMML distribution clusters most heavily in bins **0**, **2**, and **12**, which together account for **86.1%** of all tracks. Using **NOMML = 12** as the expressive criterion, the paper states that **69%** of MIDI tracks in GigaMIDI are non-expressive and **31%** are expressive. The curated expressive subset contains **869,513 files** and **1,655,649 tracks**, with **81.59% single-track files**, **18.41% multi-track files**, **28.18% drum tracks**, and **71.82% non-drum tracks**. The abstract describes this as the most substantial expressive MIDI dataset created by the authors’ heuristic.

## 6. Position within symbolic-music dataset research

GigaMIDI occupies a different niche from several prominent symbolic-music resources. **GiantMIDI-Piano** is a transcription-based dataset for classical solo piano music containing **38,700,838 transcribed notes** and **10,855 unique solo piano works composed by 2,786 composers**, with a curated subset of **7,236 works composed by 1,787 composers** [2010.07061]. **MidiCaps** is a captioned symbolic dataset built from the **Lakh MIDI Dataset**, with **168,407 MIDI files with matching captions**, and is presented as “the first openly available large-scale MIDI dataset with text captions” [2406.02255]. **MIDI-LLaMA** does not use GigaMIDI directly; instead it constructs a symbolic music–text dataset from **GiantMIDI-Piano**, described there as containing **10,855 MIDI files from 2,786 composers**, of which **9,803 musical pieces were assigned valid tags** [2601.21740].

These contrasts are structurally important. GiantMIDI-Piano is specialized in classical solo piano and derives MIDI through audio transcription; MidiCaps adds natural-language captions to a large open MIDI corpus; MIDI-LLaMA turns a piano corpus into instruction-tuning data. GigaMIDI, by contrast, is defined by multi-instrument symbolic scale, GM standardization, track-level granularity, and explicit heuristics for expressive-performance detection. This suggests that its main research role is not captioning or repertoire-specific transcription, but large-scale symbolic aggregation and filtering.

The paper’s stated downstream relevance is correspondingly broad. For MIR, GigaMIDI is presented as a consolidated corpus large enough for empirical study and feature mining. For computational musicology, it supports large-scale symbolic analysis across many instruments and styles, including contemporary styles underrepresented in notation-based corpora. For generative music, the paper argues that filtering out non-expressive MIDI may improve consistency and reduce training time for symbolic generation models.

## 7. Limitations, biases, and methodological caveats

The paper is explicit about several limitations of both the corpus and the expressive-performance heuristics [2502.17726]. The strongest compositional bias is the overrepresentation of **piano and drums**, both in GigaMIDI overall and in the curated expressive subset. The authors relate this to broader patterns of MIDI culture and digital production practice, where keyboards and drum pads are common input devices.

A second limitation concerns the MIDI standard itself. The paper states that MIDI lacks an inherent representation of hierarchical metric structure such as strong and weak beats, and that its reliance on fixed temporal grids can oversimplify expressive timing such as rubato. This is especially relevant because the heuristics operate on onset locations relative to grids, not on richer score-informed models of phrasing or performance practice.

Ground-truth coverage is also limited. Although the heuristics are intended to generalize across all GM instruments, the paper acknowledges that most reliable labeled data used in the study comes from **piano and drums**. A plausible implication is that cross-instrument generalization is argued structurally rather than validated evenly across the GM taxonomy. The paper further notes that MIDI files do not inherently distinguish between score-like, mechanically entered tracks and human-performed tracks, which is precisely the ambiguity the heuristics attempt to resolve.

Finally, GigaMIDI’s legal and technical accessibility are conditioned rather than unrestricted. It is distributed for research and private study under fair dealing, filenames are anonymized, and access is mediated through agreement on the Hugging Face Hub. This places GigaMIDI in a category of large research corpora that are designed to be usable, but not frictionless in the manner of fully open redistribution.

Source: https://www.emergentmind.com/topics/gigamidi