MAESTRO's licence covers piano recordings, not other instruments

The system

Training a model to generate expressive piano performance requires more than sheet music — it requires knowing how hard a key was struck and how long a pedal was held. Google Magenta's MAESTRO dataset page, retrieved 16 September 2026, describes exactly that: MIDI and Audio Edited for Synchronous TRacks and Organization, built from recordings of the International Piano-e-Competition. Competitors perform on Yamaha Disklaviers, acoustic grand pianos with an integrated MIDI capture and playback system, which is how the competition itself judges remote audition rounds. Google publishes the resulting dataset through Magenta, its music-and-art research group.

What the documents establish

The dataset page states MAESTRO version 3.0.0 contains 1,276 performances totalling 198.7 hours and about 7.04 million notes, split into training, validation and test sets constructed so that no composition appears in more than one split even when several contestants performed it. Audio and MIDI are aligned to roughly 3 millisecond accuracy, audio is CD quality or higher, and MIDI data preserves key-strike velocity and sustain, sostenuto and una corda pedal positions. The accompanying paper, submitted 29 October 2018, frames the release around a method it calls Wave2Midi2Wave, using notes as an intermediate representation to model musical structure across timescales the authors describe as spanning six orders of magnitude. The dataset page states the data is made available by Google LLC under a Creative Commons Attribution Non-Commercial Share-Alike 4.0 licence, and repertoire is described as mostly classical works from the 17th to early 20th century.

Craft and rights

MAESTRO documents piano only — key velocity, pedal position and audio alignment are all specific to how a Disklavier captures a piano performance, and nothing in either document extends that coverage to any other instrument. The licence matters practically: a non-commercial share-alike term means a model card or paper that lists MAESTRO among its training sources is disclosing a dataset Google itself has not cleared for commercial redistribution in its original form. That does not automatically make a downstream generative model non-commercial — licence inheritance depends on how a given model's training is structured and what its own terms say — but it is, editorially, a fact a rights-conscious reader should check for rather than assume settled whenever "trained partly on MAESTRO" appears in a system's documentation.

Outcomes and open questions

The dataset page also records that version 3.0.0 removed six recordings that had unexpected string-quartet accompaniment, present in earlier versions, and that the version 2.0.0 and 3.0.0 splits are not compatible with each other or with version 1.0.0. That matters for anyone comparing benchmark results across papers that cite different MAESTRO versions without naming one.

• Which MAESTRO version does a given piano-generation paper or model card actually cite, and are its splits compatible with the comparison being made?
• Does a commercial product's training-data disclosure address the non-commercial share-alike term, or omit it?
• Would the same alignment method generalise to other acoustic instruments with integrated capture systems, or does it depend on Disklavier-specific hardware?

MAESTRO is a well-documented, narrow resource: precise piano data under a specific non-commercial licence, not a general description of how any AI system handles piano.