How Accurate Is AI Piano Transcription? The Real Test

Every piano transcription tool describes itself as accurate, and none of the ones we checked says how it measured that. There is a standard way to measure it, used in the research these tools grow out of, and it produces useful numbers. It also measures something narrower than "will this sheet music be right", which is what you actually want to know. This page explains the metric, shows what a published model scored, and gives a test you can run on your own playing.

What the published numbers say

The figures below are from the paper that introduced MAESTRO (Hawthorne et al., ICLR 2019), computed with the mir_eval library as the paper states. They are a research model's scores, not ours and not any commercial product's.

Three things in that table matter more than the headline.

Endings are much harder than starts. On the same MAESTRO recordings, requiring the note's end to line up cost about 15 points of F1. A note starting at the right time on the right key is the easy part.

Recording conditions move the number. MAPS, described in the paper as full-length recordings from its "ENSTDkCl and ENSTDkAm" sets, is a different dataset from MAESTRO. The same model lost about 9 points of note F1 there. The paper also reports that audio augmentation during training, which includes random reverb, EQ and noise, helped on MAPS "likely because the audio augmentation made the model more robust to differences in recording environment and piano qualities".

Precision and recall differ. On the MAESTRO test split, note precision was 98.27% and recall 92.61%. The model rarely added notes that weren't there; it more often missed notes that were.

What "accurate" means in these numbers

The research standard comes from mir_eval, an open-source evaluation library. Its transcription module treats every note as a start time, an end time and a pitch, the way a piano roll does, and compares the tool's notes with a reference.

By default, following the MIREX note-tracking task it cites, a note is correct if:

• its onset is within ±50 ms of a reference note,
• its pitch is within ±50 cents (a quarter tone),
• and, in the stricter version, its offset is within 20% of the reference note's duration or 50 ms, whichever is larger.

From the matches it computes precision (how many of the tool's notes are right), recall (how many of the real notes the tool found) and F-measure, usually called F1, which balances the two. A separate module, transcription_velocity, also checks that loudness is right after rescaling.

A worked example, as arithmetic. Say a passage has 200 notes, the tool outputs 190, and 171 of them match. Precision is 171 ÷ 190 = 90%. Recall is 171 ÷ 200 = 85.5%. F1 is 2 × 171 ÷ (190 + 200) = 87.7%. That's 29 real notes missed and 19 wrong notes added, in 200.

Scale it up: at the MAESTRO test recall of 92.61%, a piece with 1,000 notes has about 74 not matched. Most readers would call 95% F1 excellent, and it is, for a machine. It still means corrections.

What the percentage leaves out

The metric compares notes as sounds in time. Sheet music is a set of decisions about how to write those sounds down, and none of them is scored.

• Note spelling. G sharp and A flat are the same key and the same pitch, so both count as correct.
• Rhythmic notation. A note that starts 30 ms late is correct by the metric, but notation software has to decide whether it's on the beat or a sixteenth after it. Rubato makes that decision harder.
• Hand splitting and voices. Which staff a note goes on, and which stem direction, isn't part of the comparison.
• Key and time signature, bar lines, pedal marks. Not measured.

So two transcriptions with the same F1 can be very different to read. That is why we care more about the MusicXML than the PDF: it's the file you fix these things in, in MuseScore, Sibelius or Dorico. For what our own tool produces from a recording, our comparison of audio-to-MIDI converters sets the file types side by side.

Why vendors don't give you numbers

We checked the product pages of the best-known tools on 16 September 2026. Klangio's Piano2Notes FAQ says it "delivers highly accurate results — especially with clear audio". Songscription's FAQ says its technology "is incredibly accurate for single-instrument recordings, particularly piano", and that for other recordings "the accuracy may vary". Spotify's Basic Pitch README says it "works best on one instrument at a time". Human services are no different: Music Notation Hub's transcription page advertises a "99% Accuracy Rate" without saying how it's counted.

None of those is dishonest; all of them are unmeasurable as written. Our own FAQ says the result "depends on the recording": solo piano recorded close in a quiet room gives the cleanest scores, and dense pedalling, heavy reverb or other instruments can need corrections. We don't publish an accuracy percentage for Piano to Sheet Music, and our editorial rule is not to make accuracy claims without a stated method behind them. The protocol below is the method any such figure would have to come with — and we have since run a version of it on our own pipeline and published the per-piece numbers in how we measure piano transcription accuracy.

The failure cases, and where they come from

Sustain pedal. MAESTRO's MIDI records sustain pedal positions separately from key presses. With the pedal down, a string keeps sounding after the key comes up, so the audio gives no clear end to the note. The paper doesn't break its errors down by cause, but the 15-point gap between onset-only and offset F1 above is consistent with endings being the ambiguous part. In notation it appears as long ties and overlapping notes. We cover the fixes in sustain pedal transcription errors.

Room and microphone. A phone across a room records reflections as well as strings. The MAPS drop in the table is the research version of that problem.

Other instruments. Basic Pitch's README says it "works best on one instrument at a time", and Songscription's FAQ says accuracy "may vary" for recordings with many other instruments. A singer, a guitar or a backing track is sound a piano tool has to separate from the piano first.

Tuning. mir_eval allows ±50 cents because pitches are compared to a tuned grid; its documentation gives "a semitone grid tuned to 440 Hz" as the example. A piano far enough from standard pitch sits between grid lines. Our guide to transcribing an out-of-tune piano covers what that does.

Repertoire. MAESTRO's repertoire is, in its own words, "mostly classical, including composers from the 17th to early 20th century". A benchmark score says less about a jazz voicing or a pop ballad than about a Chopin étude.

A test you can run on your own recording

This is the protocol, at two levels of effort. It judges the tool on the only recording that matters: yours.

The quick version, no code

1. Pick the hardest 20 to 30 seconds of the piece: the pedalled climax, the fast run, the thick left hand. Not the intro.
2. Record it the way you'll record the whole piece, same phone, same room.
3. Run the identical file through each tool's free preview. Piano to Sheet Music transcribes the first 30 seconds free, and other tools have similar demos.
4. Open the MusicXML, not the PDF, in MuseScore Studio.
5. Count the fixes per tool in three columns: wrong or missing pitches, rhythm fixes, and notes on the wrong staff.
6. Pick the tool with the fewest fixes on your passage, not the best marketing.

If you don't read fluently, you can still do step 5 by listening: MuseScore's playback against your recording shows where they disagree. Our guide to proofreading a score by ear walks through it, and MuseScore's own audio-to-sheet-music options explains what MuseScore itself can and can't import.

The measured version, with a digital piano

A digital piano gives you what the researchers had: a reference. Record the same take two ways at once, MIDI from the piano's USB port and audio from a phone or the piano's audio output. The MIDI is ground truth for what you played.

1. Transcribe the audio with each tool and export MIDI.
2. Load both MIDI files as note lists of onset, offset and pitch.
3. Score them with mir_eval, once on onsets and pitch only, once with offsets. Use the defaults so your numbers mean the same thing as the paper's.
4. Repeat under different conditions: close and dry, pedal-heavy, across the room.
5. Then do the quick-version fix count on the MusicXML, because the F1 won't tell you about readability.

Two caveats. Timing between the MIDI and audio must line up, or every onset misses its 50 ms window; trim both to start at the same first note. And a digital piano's sound isn't an acoustic piano's, so a good result here is evidence about clean input, not about your upright in the living room.

So, how accurate is it?

Accurate enough on clean solo piano to save most of the writing, and not accurate enough to skip checking. The research figures say onsets and pitches are largely solved on good recordings, note lengths less so, and that recording conditions matter a great deal. Whether a tool is worth paying for is a question about your recording and your tolerance for editing, which is also the honest line in transcribing by ear vs AI transcription. If you want to try ours on your hardest passage, you can try it here.

Frequently asked questions

How accurate is AI piano transcription?

On a clean benchmark, very. Google's Onsets and Frames model trained on MAESTRO scored a note F1 of 95.32% on the MAESTRO test set, counting a note correct if it starts within 50 ms at the right pitch. When note endings also had to match, F1 fell to 80.50%. Your own recording can score quite differently.

What does an accuracy percentage for transcription mean?

Usually nothing specific unless the method is stated. Researchers use note precision, recall and F1 from the mir_eval library: a note counts as correct if its onset is within 50 ms and its pitch within 50 cents of the reference, optionally with a matching offset. A percentage without that definition and a named test set can't be compared.

Why does transcription do worse on my recording than the benchmarks?

Benchmarks are recorded in ideal conditions. MAESTRO's audio comes from Yamaha Disklavier concert grands at CD quality or higher. When the same MAESTRO-trained model was tested on MAPS, a dataset with different pianos and recording setups, its note F1 dropped from 95.32% to 86.44%. Phones, rooms and pedal all move you away from benchmark conditions.

Does a high note accuracy mean the sheet music will be readable?

No. The standard metrics compare notes as start time, end time and pitch, like a piano roll. They don't check note spelling, hand splitting, rhythmic notation, voices or key signature, which decide whether a score is readable. A draft can match most notes and still need editing before you can sight-read it.

How can I test a transcription tool's accuracy myself?

Use the hardest 20 to 30 seconds of your own piece in each tool's free preview, not the easy intro. If you own a digital piano, record MIDI and audio of the same take: the MIDI is your reference. Then count what you would have to fix in the MusicXML: wrong pitches, rhythms and hand splits.

What is the hardest thing for piano transcription to get right?

Note endings. In the MAESTRO paper, the model's note F1 was 95.32% on onsets and pitch alone but 80.50% once offsets had to match. Sustain pedal, which keeps strings ringing after the key is released, is one reason endings are ambiguous. Room echo and other instruments in the mix add further errors.

Sources

• Hawthorne et al., "Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset", ICLR 2019
• Google Magenta: The MAESTRO Dataset
• mir_eval documentation: mir_eval.transcription
• mir_eval documentation: mir_eval.transcription_velocity
• Klangio: Piano2Notes
• Songscription
• Spotify Basic Pitch repository
• Music Notation Hub: Music transcription

Vendor pages checked 16 September 2026.