Gangqinpu: 1,000 hours of piano with the score attached, and a licence that keeps it in the lab

Three researchers, Wei Zeng, Anvay Mathur and Ye Wang, have posted a paper on arXiv introducing Gangqinpu, a piano dataset that lines up three things at once: a recording, a MIDI version of the performance, and the written score. It was submitted on 26 October.

What it is, plainly

Most music software that "listens" to piano has to learn from examples where someone already knows the right answer. Gangqinpu supplies 18,786 such examples, score-performance pairs totalling 1,007.1 hours of source audio, with every note in the performance matched to a note in the score and beat and downbeat markers derived from that match (arXiv).

It covers pop, classical, jazz, folk/traditional, rock and blues. The authors say datasets that align all three formats are scarce and limited in scale and genre (arXiv). The best-known one, ASAP, has audio for only 519 of its 1,067 performances and is restricted to Western classical piano (paper, full text).

The whole thing is built automatically from online audio-score pairs. A transcription model (Transkun v2) turns the audio into MIDI, that MIDI is aligned to the score, an AI model assigns genre labels, and a small classifier throws out "deadpan" renderings that were just synthesised straight from notation. On 288 expert-labelled test samples that classifier scored 98.61% accuracy (paper, full text). The authors also split the data jointly with ASAP so the same piece doesn't leak into both training and test sets.

The licence

This is the part to read twice. Per the project README, the code is MIT, but the dataset and classifier weights are CC BY-NC-SA 4.0: non-commercial use only, and you have to confirm that on the Hugging Face page before downloading. The audio and original scores are not redistributed at all; you get source URLs and a download script, the README warns that links may expire, and it says the underlying works stay subject to their rights holders' terms (GitHub).

Why it matters for transcription tools

The paper tested Gangqinpu as extra training data on three jobs: beat tracking, turning performance MIDI into a score, and "expressive rendering" (making quantised MIDI sound played). Beat tracking improved across datasets; gains for transcription and rendering were mostly in-domain, meaning on Gangqinpu-style material (arXiv). The authors explicitly point at DAW use: turning quantised MIDI into natural-sounding performances (paper, full text).

What we think

The genre spread is the real news. Anyone who has run a pop ballad or a jazz comp through an audio-to-score tool knows the output often reads like a classical exam piece gone wrong, and a training diet of mostly classical repertoire is a plausible reason. More pop and jazz data aimed at that gap is welcome. That diagnosis is ours, not the paper's.

But temper the excitement. The performance MIDI is itself transcribed by a model, so anything trained on it inherits Transkun's mistakes; this is machine-labelled data, not ground truth from a sensor-equipped piano. And "in-domain" gains mean you shouldn't expect your notation app to get better at your playing overnight.

The licence decides who benefits. Non-commercial plus share-alike means university labs and hobby projects can use it freely, while a company shipping a paid transcription plug-in can't train on it without a separate deal, and it still couldn't clear the recordings, which Gangqinpu doesn't own. We think that is the honest setup: publishing links rather than audio is a sensible line on data built from other people's uploads. The method and MIT code are the more portable gift: a tool maker with licensed recordings can rebuild the same pipeline on their own catalogue.