GiantMIDI-Piano: Large-Scale Solo Piano MIDI Dataset

• GiantMIDI-Piano is a large-scale symbolic dataset offering classical solo piano recordings transcribed into MIDI with detailed composer and work metadata.
• It is constructed via an automated pipeline that retrieves YouTube audio, applies a robust solo piano detector, and uses a high-resolution CRNN transcription system.
• The dataset supports applications in music information retrieval, composer classification, and computational musicology, despite omitting full score annotations.

GiantMIDI-Piano is a large-scale symbolic dataset for classical solo piano music created by transcribing internet audio recordings into MIDI and organizing them with composer- and work-level metadata extracted from the International Music Score Library Project (IMSLP). Introduced by Kong et al. as a resource for music information retrieval, piano transcription, symbolic music generation, computational musicology, and expressive performance analysis, it contains 38,700,838 transcribed notes, 10,855 unique solo piano works, 2,786 composers, and 1,237 hours of music; a curated subset contains 7,236 works, 1,787 composers, 24,253,495 notes, and 875 hours (Kong et al., 2020).

1. Scope, identity, and corpus composition

GiantMIDI-Piano addresses what Kong et al. describe as a lack of large-scale symbolic datasets for classical piano music. The dataset is explicitly positioned as a classical solo piano corpus in MIDI format rather than a mixed-instrument symbolic archive, and this specialization is central to both its construction and its intended uses (Kong et al., 2020).

The full release contains 10,855 unique solo piano works composed by 2,786 composers and totals 38,700,838 transcribed notes across 1,237 hours. The paper also reports a curated subset with stricter metadata filtering, containing 7,236 works by 1,787 composers, totaling 24,253,495 notes across 875 hours. In the full dataset, 90% live-performance MIDI files and 10% sequence-input MIDI files are reported; the curated subset is reported as 89% performed MIDI (Kong et al., 2020).

The dataset is described in the paper as the largest piano MIDI dataset so far. Its design distinguishes it from aligned performance corpora such as MAESTRO: GiantMIDI-Piano is not based on paired studio recordings and score-aligned MIDI capture, but on a pipeline that begins with bibliographic metadata, retrieves audio from the web, filters for solo piano, and then applies automatic transcription (Kong et al., 2020).

This scale has made GiantMIDI-Piano relevant beyond its original release. A later composer-classification study uses it as a transcription-based dataset for large-scale MIDI-based composer classification and reports experiments with up to 100 composers (Kong et al., 2020). A later multimodal LLM study, MIDI-LLaMA, uses GiantMIDI-Piano as the basis for a symbolic music–text corpus and states that the dataset contains 10,855 MIDI files from 2,786 composers (Yang et al., 29 Jan 2026).

2. Data acquisition pipeline

The construction pipeline begins with IMSLP, from which the authors parsed composer pages and work lists. They report extracting 143,701 music works and 18,067 composers. For composers, they also attempted to extract birth year, death year, and nationality from biography links; for a smaller set of 2,786 composers, they manually labeled 2,291 birth years, 2,254 death years, and 2,115 nationalities, with unknown values retained as "unknown" (Kong et al., 2020).

Audio retrieval proceeds by querying YouTube for each IMSLP work using a keyword composed of the composer first name, composer surname, and work name. The pipeline selects the first returned result for each query, producing 60,724 audio recordings downloaded from the 143,701 works (Kong et al., 2020).

To decide whether a returned title matches the intended work, the paper uses a modified Jaccard similarity. If $X$ is the set of words in the search keyword and $Y$ is the set of words in the returned YouTube title, the conventional Jaccard index is described as

$J = \frac{|X \cap Y|}{|X \cup Y|}$

but the paper instead defines

$J = \frac{|X \cap Y|}{|X|}$

and applies a threshold of 0.6, so that if $J > 0.6$, the keyword and title are considered matched (Kong et al., 2020).

Because the retrieved recordings are not guaranteed to be solo piano, the authors train an audio-based solo piano detector. The detector operates on 1-second audio segments, uses log mel spectrograms as features, and applies a CNN with four convolutional layers. Each convolutional layer uses $3 \times 3$ convolution, batch normalization, and ReLU, and the detector is trained with binary cross-entropy. At inference, segment-level predictions are averaged, and an audio recording is classified as solo piano if the probability is > 0.5 (Kong et al., 2020).

From the 60,724 downloaded recordings, the detector identifies 10,855 solo piano recordings by 2,786 composers. These recordings are then transcribed into MIDI and constitute the full GiantMIDI-Piano dataset (Kong et al., 2020).

To improve metadata precision, the authors create a curated subset by requiring that the downloaded audio title contain the surname of the composer. This surname constraint yields 7,236 works by 1,787 composers, and the metadata includes a youtube_title_contains_surname flag (Kong et al., 2020).

3. Symbolic representation and transcription system

Each GiantMIDI-Piano file is a transcription-derived MIDI representation rather than a score-native encoding. The paper states that each transcribed MIDI file contains onset, offset, pitch, and velocity attributes of piano notes, together with pedals represented as sustain-pedal onset/offset events (Kong et al., 2020). This note- and control-level structure is one reason later work treats GiantMIDI-Piano as well suited to symbolic music understanding and composer classification (Kong et al., 2020, Yang et al., 29 Jan 2026).

The transcription backend is the authors’ open-source high-resolution piano transcription system, trained on the MAESTRO v2.0.0 training set. The model uses a CRNN with 8 convolutional layers and 2 bi-directional GRU layers, and has separate output modules for frame-wise pitch classification, onset regression, offset regression, and velocity regression. The output dimension is 88, corresponding to the 88 piano keys. For pedal transcription, the model is similar except that it uses a single output for pedal onset/offset (Kong et al., 2020).

The inference front end converts audio to mono at 16 kHz. The STFT uses a Hann window size = 2048 and hop size = 160, giving a time resolution of 100 frames per second. The mel front end uses 229 mel bins, and the model consumes log mel spectrograms. Predicted outputs are then post-processed into MIDI events (Kong et al., 2020).

A limitation explicitly noted in the paper is that the dataset does not provide pitch spellings distinguishing enharmonics, beats, time signatures, key signatures, or scores, and it does not disentangle score content from performance expression (Kong et al., 2020). These omissions are important when interpreting the symbolic content: GiantMIDI-Piano is a rich note-event corpus, but not a fully notated score corpus.

4. Evaluation, quality control, and error characterization

The release evaluates three components: solo piano detection, metadata accuracy, and transcription quality (Kong et al., 2020).

For solo piano detection, the authors manually labeled 200 randomly selected music works from the 60,724 downloaded recordings. At threshold 0.5, the detector achieved Precision: 89.66%, Recall: 86.67%, and F1: 88.14% (Kong et al., 2020).

For metadata accuracy, the authors manually checked 200 solo piano works from the full dataset. They report that 174 / 200 were correctly matched, corresponding to Metadata accuracy: 87%. After applying the surname constraint, 136 / 140 were correctly matched, corresponding to Precision: 97.14%. In the 200-example sample, the paper reports 180 live performances and 20 sequenced MIDI files (Kong et al., 2020).

The paper also provides composer-specific retrieval accuracies before and after surname filtering.

Composer	Before surname filtering	After surname constraint
Bach	59%	75%
Mozart	71%	82%
Beethoven	54%	94%
Chopin	37%	82%
Liszt	90%	96%
Debussy	76%	90%

These figures show that metadata precision varies substantially by composer and that surname filtering improves precision markedly (Kong et al., 2020).

The transcription system itself is evaluated on the MAESTRO test set. The reported metrics are Onset F1: 96.72%, Onset + offset F1: 82.47%, and Onset + offset + velocity F1: 80.92%. For pedals, the reported metrics are Pedal onset F1: 91.86% and Pedal onset + offset F1: 86.58%. The paper states that this system beats the prior Onsets and Frames system in onset F1, comparing 94.80% to 96.72% (Kong et al., 2020).

Because GiantMIDI-Piano lacks ground-truth aligned reference MIDI for its recordings, the authors assess quality on the dataset itself through an alignment error rate (ER):

$ER = \frac{S + D + I}{N}$

where $S$ denotes substitutions, $D$ deletions, $I$ insertions, and $Y$0 the number of reference notes (Kong et al., 2020).

To interpret GiantMIDI-Piano’s error, they write

$Y$1

and, for MAESTRO,

$Y$2

They then define a relative error,

$Y$3

as a rough estimate of the transcription error contribution in GiantMIDI-Piano (Kong et al., 2020).

On 52 shared works, the paper reports MAESTRO median ER: 0.061, GiantMIDI-Piano median ER: 0.154, and Relative median ER: 0.094. Median component rates are also reported: for MAESTRO, D = 0.009, I = 0.024, S = 0.018, ER = 0.061; for GiantMIDI-Piano, D = 0.015, I = 0.051, S = 0.069, ER = 0.154; and for the relative difference, D = 0.006, I = 0.026, S = 0.047, ER = 0.094. The paper notes that insertions are more common than deletions in the relative comparison (Kong et al., 2020).

5. Musical analysis and computational use cases

A central claim of the release is that GiantMIDI-Piano is large enough for composer- and era-level musical analysis. To demonstrate this, the paper analyzes works by J. S. Bach, Mozart, Beethoven, Chopin, Liszt, and Debussy using pitch-class, interval, trichord, and tetrachord statistics (Kong et al., 2020).

For pitch classes, the paper uses the twelve classes

$Y$4

mapped to 0–11. Reported findings include: Bach showing strong use of D, E, G, A; Mozart emphasizing C, D, F, G and using more A♯/B♭; Beethoven favoring C, D, G; Chopin favoring D♯/E♭ and G♯/A♭; and Liszt and Debussy using pitch classes more uniformly. The paper further states that Romantic and later composers tend to explore more distant keys than Baroque/classical composers (Kong et al., 2020).

For intervals, the paper defines

$Y$5

where $Y$6 is the MIDI pitch of note $Y$7. It considers both melodic intervals between successive notes and harmonic intervals between notes sounding at the same time, examining ordered intervals from -11 to 11. The paper reports that all six composers use major second and minor third most, that intervals are not symmetric around 0, that Bach and Mozart use more downward major seconds than upward ones, that repeated notes (interval 0) are less common in Bach, that major seventh and tritone are least used across composers, and that Romantic/later composers use intervals more uniformly (Kong et al., 2020).

For trichords, notes within a 50 ms window are grouped as chords; windows are non-overlapping; notes can belong to only one chord; only 3-note sets are kept; chords are transposed to C-based prime forms; and equivalent chords are merged. Examples given are the major triad $Y$8 and minor triad $Y$9. The paper reports that all composers use the major triad $J = \frac{|X \cap Y|}{|X \cup Y|}$0 most, the minor triad $J = \frac{|X \cap Y|}{|X \cup Y|}$1 is usually second, Liszt uses more augmented triads $J = \frac{|X \cap Y|}{|X \cup Y|}$2, and Debussy uses more $J = \frac{|X \cap Y|}{|X \cup Y|}$3 and $J = \frac{|X \cap Y|}{|X \cup Y|}$4 (Kong et al., 2020).

For tetrachords, the same 50 ms chord windowing is used, but only 4-note sets are retained and chords are transposed to root position. Examples include the dominant seventh $J = \frac{|X \cap Y|}{|X \cup Y|}$5, diminished seventh $J = \frac{|X \cap Y|}{|X \cup Y|}$6, and minor seventh $J = \frac{|X \cap Y|}{|X \cup Y|}$7. The paper reports that Bach, Beethoven, Mozart, and Chopin use dominant seventh most, Liszt uses diminished seventh most, and Debussy uses minor seventh most; it also states that Debussy’s tetrachord distribution is notably different from the others (Kong et al., 2020).

The paper presents these analyses as evidence that the dataset can support computational musicology. A plausible implication is that the corpus scale is sufficient to stabilize style statistics for many composers, although the release itself demonstrates this only for the six-composer case study.

6. Downstream adoption, related work, and limitations

GiantMIDI-Piano has been used directly in later symbolic-music research. In large-scale MIDI-based composer classification, it serves as the underlying transcription-based dataset for systems that use piano rolls, onset rolls, and velocity rolls as input representations and deep neural networks as classifiers. Using convolutional recurrent neural networks, that study reports 10-composer and 100-composer classification accuracies of 0.648 and 0.385 on 30-second clips, and 0.739 and 0.489 on music pieces, respectively; it also states that the MIDI-based system outperforms several audio-based baseline classification systems (Kong et al., 2020).

In multimodal symbolic music understanding, GiantMIDI-Piano is the basis for MIDI-LLaMA’s music–text corpus. That work states that GiantMIDI-Piano contains 10,855 MIDI files from 2,786 composers and uses it as a source of classical piano performances paired with language supervision. It builds a fine-grained annotation pipeline from each piece’s title and composer, retrieved contextual text from classical-music websites, GPT-4o extraction of genre, style, composition background, expressive intent, and perceived emotion, and manually verified outputs. In a stratified sample of 100 works, two experts judged annotations acceptable at 89% acceptance for categorical tags and 93% for descriptive annotations. The corpus is further expanded by extracting three non-overlapping 20-second clips from long pieces, aligned to bar lines when possible, yielding 29,409 clips and about 2.3 million Q/A pairs (Yang et al., 29 Jan 2026).

The dataset also sits within a broader ecosystem of piano and MIDI resources. A large-scale MIDI-to-sheet retrieval study does not mention GiantMIDI-Piano explicitly, but it is relevant in spirit because it addresses scalable methods for linking large symbolic piano corpora to IMSLP sheet music via modified symbolic bootleg scores and reverse-index retrieval (Tsai, 2020). A later large-scale symbolic generation study cites GiantMIDI-Piano as one of the available piano-MIDI datasets, although it does not use it experimentally (Lehmkuhl et al., 10 Nov 2025). A large general-MIDI corpus study, GigaMIDI, explicitly references GiantMIDI-Piano as a dataset of AI-transcribed piano tracks that may contain expressive performance nuances, but excludes it from that study’s own expressive-performance ground-truth curation (Lee et al., 24 Feb 2025).

Several misconceptions are clarified by the primary release. GiantMIDI-Piano is not a score-aligned, notation-complete corpus; it is a transcription-derived MIDI dataset without beats, time signatures, key signatures, or scores (Kong et al., 2020). It is also not restricted to manually verified metadata throughout the entire full release; instead, the curated subset provides the stricter precision-oriented variant through the composer-surname constraint (Kong et al., 2020). Finally, although the data contain expressive attributes such as velocity and pedals, the paper explicitly notes that the dataset does not disentangle score content from performance expression, so analytical use requires care when separating compositional regularities from performance-derived variation (Kong et al., 2020).