What AI Changed in Music: The Separating Side, Not the Generating Side

Why I looked into this

I follow AI news daily. Protein structures got solved, open problems in mathematics started cracking, code gets reviewed before it’s written.

So what about piano? What about music?

As someone who plays Hanon every morning, I was genuinely curious — because I hadn’t felt much change around my own practice. It turns out there has been plenty of change. It’s just that the place everyone is looking is not the place where it landed.

In generative AI, what happened was institutional, not technical

Start with the loud part. Suno and Udio generate lyrics, vocals, backing, and mix in one shot, at commercial quality. Suno is on the v5 line as of 2026.

But what actually happened in this space since 2024 was less a technical breakthrough than a legal settlement.

When	What
June 2024	The three majors (UMG, WMG, Sony Music) sue Suno and Udio for copyright infringement
October 2025	UMG settles with Udio and partners
November 2025	WMG settles with and partners with both Suno and Udio
2026	Migration to models trained on licensed catalogs. Sony’s claim is still live

The question moved from “can it be made?” to “who counts as having made it, and who gets paid?” In early 2026 there were reports of Suno changing its terms around ownership of outputs (Suno denied any change) — the most basic question of all, who owns the song, is still unsettled.

This is not an AlphaFold-shaped revolution where something unknown got solved. It’s a question of how the economics of mass production get rearranged. That matters, but it’s a different kind of thing.

What actually landed was the side that takes music apart

This is where the ground moved. Things that were impossible ten years ago are now routine.

Source separation

Pulling vocals, drums, bass, and everything else back out of a finished mix. Spleeter (Deezer, 2019) started the wave by going open source; Meta’s Demucs, particularly Hybrid Transformer Demucs (2022), became the de facto standard. Benchmark efforts like Sony’s MDX Challenge accelerated the quality curve.

The emblematic case is the Beatles’ Now and Then (2023). From a degraded 1970s demo cassette, Peter Jackson’s team isolated John Lennon’s voice well enough to finish the song. The technique was developed during Get Back (2021), and the track won Best Rock Performance at the 2025 Grammys.

For a player, the implication is blunt: you can build a minus-one practice track out of any recording. Just the orchestra of a concerto. Just the other part of a duet.

Automatic transcription

Going from performance audio to MIDI, and then to readable notation. The lineage:

• Onsets and Frames (Google Magenta, 2018) — separating onset detection as its own objective made accuracy jump
• MAESTRO (2019) — roughly 200 hours of audio with aligned MIDI; the training foundation
• ByteDance high-resolution piano transcription (2020) — high accuracy including pedal
• Nakamura et al., Kyoto University (2021) — reported as the first system producing notation at a practically usable level
• MT3 (Google, 2021) — multi-task, multitrack

By 2026 there are enough working audio-to-score tools that comparison articles exist. What hasn’t changed is that the output is rarely usable as-is; human correction is still assumed.

OMR (optical music recognition)

Turning printed scores and manuscripts into machine-readable data. Unglamorous, but everything in the musicology section below rides on it.

Put the three together and the result is one sentence: the entire recorded and published stock of music became data.

Why piano specifically works

The most useful distinction in music AI is between the symbolic and audio domains.

Domain	Representation	Typical tasks
Symbolic	Score, MIDI, MusicXML	Composition, arrangement, fingering, style analysis
Audio	Waveform, spectrogram	Generation, separation, timbre transfer

Transcription (audio → symbolic) and performance rendering (symbolic → audio) are the bridges, and that’s where the hard problems sit. Structurally it mirrors ASR and TTS in speech AI, but music is polyphonic, which makes it a tier harder: you have to hear several simultaneous notes apart.

Piano does well for three reasons. Pitch is discrete and the tuning holds. Large volumes of performance data with aligned MIDI exist (Disklavier and friends). And once a string is struck, the pitch doesn’t move — it only decays.

The flip side: strings and voice, where pitch moves continuously through vibrato and portamento, remain the wall. Pianists happen to play the instrument machines find easiest.

Where “solving the unknown” actually stands

The closest thing to genuine discovery is computational musicology. In June 2026, the National Library of Norway and the University of Oslo hosted DAIM 2026 (Symposium on Digital and AI Musicology), themed around preservation, transcription, analysis, discovery, and interconnection — a good snapshot of where the field’s attention is.

What’s moving:

• Attribution — authenticity judgments from statistical analysis of handwriting and style, previously dependent on a handful of experts’ ears and instincts
• Quantified performance style — analyzing historical recordings at scale to trace how tempo, rubato, and dynamics actually shifted, rather than arguing from impression
• Mass transcription — manuscripts that nobody had read becoming searchable

Honestly though: as of 2026 there is no single big discovery you could point to and say AI solved a long-standing mystery. What exists is the infrastructure for exhaustive survey. There are headline projects like the AI completion of Beethoven’s Tenth (2021), but that’s closer to composition than to discovery.

In an adjacent field, Bellier et al. (2023) reconstructed Pink Floyd’s Another Brick in the Wall recognizably from intracranial recordings of auditory cortex activity. That one does move into genuinely unknown territory.

Following the player, not replacing them

The technology that performance actually welcomes isn’t the kind that replaces humans — it’s the kind that tracks human fluctuation.

IRCAM’s Antescofo is a fixture for synchronizing electronics in contemporary music. Yamaha’s AI ensemble technology has a player piano follow a human soloist. Dear Glenn (2019), which learned Glenn Gould’s style, belongs to the same lineage.

2026 research includes real-time performance scoring via MIDI alignment running on embedded hardware, and multimodal work that fuses audio and video to correct fingering and model expressiveness at the same time. All of it assumes there’s still a person playing.

What hasn’t happened

To avoid overstating things, here’s what didn’t move.

• The value structure of performance is unchanged. Embodiment and live presence sit outside AI’s reach. If anything, as generated music saturates, the scarcity of “a person playing, here, now” has gone up in relative terms
• Instrument innovation isn’t AI-driven. Physical modeling engines like Pianoteq are simulation, not AI — a common conflation
• No new music theory has emerged. The scale at which existing theory can be tested expanded; the theory itself didn’t
• Automatic evaluation of “good playing” is immature. Technical accuracy is measurable; for musical persuasiveness there isn’t even an agreed metric

What changed for the player

In the end, AI landed on music not as “it writes new songs” but as all past music became data, and exhaustive survey became possible for the first time.

And because digital pianos emit MIDI, that reaches the practice desk. Record yourself as MIDI and timing jitter, velocity spread, and hand-to-hand offset become numbers. Align that against the score and you can pin down exactly where the tempo collapses and which notes you’re dropping. Mechanical work like Hanon and scales is the best fit of all, because deviation is easy to detect.

Pull stems with Demucs, analyze your own playing with partitura or pretty_midi. All open source, all local. It isn’t a grand discovery, but as a tool for raising the resolution of a morning practice session, it’s already well past usable.

I’ve written about the theory side before — the circle of fifths tool and the piano practice tools — but I haven’t touched performance analysis yet. That’s what I want to try next.

One more thing nagged at me after writing this. If music only moved on the separating side, what about every other field? Lining them up, the dividing line turned out not to be how hard a field is, but whether you can check the answer. That’s the follow-up: The Verifier Decides.