Harmonic analysis, drawn on the score
Ask what is happening in the music, and give your score as a MusicXML file or a PDF of printed sheet music: attach it in ChatGPT, or in Claude choose it in the card that appears. You get the engraved score with the analysis marked on it, the way a theory teacher would pencil it in, plus a short written commentary.
Chopin, Nocturne Op. 9 No. 2, bars 1–8. Tap or click to enlarge. PDF
What it marks
- The chords: each chord is labelled under the music with a Roman numeral (I, IV, V…), which says which note of the key the chord is built on. A bracket under the numerals names the chord's job: "tonic" for home, "pre-dom." (predominant) for moving away, "dominant" for pulling back home.
- Endings and phrases: cadences, the musical equivalent of commas and full stops, are boxed; brackets above the music show where each phrase, like a sentence, begins and ends.
- Notes outside the chord: melody notes that don't belong to the chord underneath, such as passing notes that connect two chord notes, are circled and named.
- How the parts move: arrows follow a note as it moves to its next note, where that movement matters, for example a tense note settling onto a calm one.
- A written explanation: numbered circles on the page match numbered paragraphs in the chat reply, which say what you might have expected at that moment and what the composer did instead.
How it works
- The score is readA MusicXML file is read as it is; a PDF is read by music-recognition software that turns the printed page back into notes.
- The chat model analysesClaude or ChatGPT works through a beat-by-beat table of every note, labels the keys, chords, cadences and phrases, and picks out what is worth pointing out to you. In writing for two to four parts, code also flags part-writing slips such as parallel fifths.
- The score is drawnThe analysis is printed onto a properly engraved copy of the score, as an image and a PDF. Nothing is drawn by an image-generating AI.
Three views of the same passage
When the full page is too busy, ask for one layer at a time: the chords (harmony), how the individual parts move (voice leading), or how the piece is built from phrases (form).
Big scores, made easier to read
An orchestra or band score can have twenty or more lines of music stacked on top of each other. For any score with more than four, you'll be offered a piano version of the same bars: the whole ensemble squeezed onto the two lines a pianist reads, with the tune on top. Small labels show which instruments are playing the tune at each point, using the usual short names: Vl. for violins, Fl. for flute, Tutti for everyone together. The analysis stays the same. Only the page gets simpler.
Mozart, Symphony No. 41 "Jupiter", first movement, bars 1–4.
How good is it?
Agrees with an expert 99% as often as experts agree with each other.
Tested chord by chord on 36 pieces, against two published expert analyses of every piece.
The pieces are Bach chorales, movements from Mozart and Beethoven piano sonatas, and Chopin mazurkas. For each chord an expert wrote, we checked whether our analysis said the same thing for more than half of that chord: the same chord, in the same key, with the same note in the bass. We did the same for the two experts against each other.
- We agreed with 69% of an expert's chords; one expert agreed with 70% of the other's. The 99% is the first figure divided by the second, before rounding. The difference is within the noise.
- Counting only the main note of each chord (for example C in a C major chord, whichever way it is turned), we agreed with 88% of an expert's chords, and the experts with 91% of each other's.
- Experts don't fully agree with each other either. Measured the same way, they agreed on 83% of the chords in Bach chorales and 79% in Mozart sonatas, but only 59% in Beethoven sonatas and 58% in Chopin mazurkas. We agreed with an expert less often than that on the chorales and Mozart (77% and 73%), and more often on Beethoven and Chopin (63% and 66%). Where the experts disagree, matching either one earns us half credit while the experts get none, which flatters the comparison a little.
These results were measured on passages, not whole pieces, with Claude Opus writing the analysis. We have not measured other chat models, including ChatGPT, the same way, and they may do less well. The expert analyses are published, so the model may have seen them. How we tested it, step by step.
So treat it like a well-read study partner, not an answer key. It is least reliable in the places where experts often disagree too: when a piece in a major key briefly borrows chords from the minor key, it sometimes says the music has moved into the minor; and music that doesn't follow the usual chord patterns, such as Debussy's, gets looser descriptions.
Get it
- Claude: add the Sheet Music Analysis plugin (steps), then ask for an analysis in a chat and choose your score file in the card that appears. A link to your score works too.
- ChatGPT: add Sheet Music Analysis from ChatGPT's plugin directory (steps), then attach your score in a chat.
It is free, needs no account, and shows no ads. See How to use for example prompts and what each mark means.
Support
Questions, wrong analyses worth a look, or problems with a file: support@sheetmusicanalysis.com.





