PIANO TO MIDI · CHORDS SUPPORTED · NO UPLOAD
Piano to MIDI: Convert a Piano Recording to Notes
Solo piano is one of the better sources for audio-to-MIDI conversion. Choose a piano recording on the homepage converter, and it detects notes — including several notes at once in chords — and gives you a .mid file to edit. Expect to clean up some extra or missing notes, especially with heavy pedal or reverb.
How to convert, step by step
- Open the converter. Go to the MP3toMIDIFlow homepage. The converter is at the top of the page; nothing needs to be installed.
- Choose your file. Select or drag in one piano recording (MP3, WAV, M4A, FLAC, OGG, AAC or WMA, or the audio in an MP4/MOV video). Solo piano with little pedal and reverb works best.
- Convert in the browser. Select Convert to MIDI and keep the tab open. The note-detection model and decoder load from this site and run on your device; your audio is not uploaded.
- Download and edit. Download the .mid file and open it in a DAW or notation editor to set the tempo, fix wrong notes and choose an instrument sound.
Why piano works comparatively well
A piano note has a clear, sharp attack and a stable pitch, which makes note starts easier to find than with a sung or bowed note. The note-detection model used here, Spotify's open-source Basic Pitch, is polyphonic: it estimates every pitch it hears at each moment, so chords and two-hand playing can produce simultaneous notes in the MIDI. Its pitch range covers the 88-key piano range.
The output is a single track with notes, timing and velocities based on loudness. Both hands end up on the same track; if you want left and right hand separately, split the track in your DAW or notation editor by pitch.
What makes piano transcription harder
- Sustain pedal. With the pedal down, earlier notes keep ringing under new ones. The model may extend those notes, add notes that are still ringing, or cut them short as they fade. You may need to fix note lengths and add pedal (CC64) yourself.
- Overtones. Low piano notes have strong harmonics, so the model sometimes adds a quieter note an octave or a twelfth above. Our extra chord notes guide shows how to check before deleting.
- Dense chords and fast runs. Thick voicings and fast passages produce more missed and merged notes.
- Room sound and recording quality. A phone recording in a reverberant room is harder than a close-miked or digital piano recording.
- Piano in a band mix. The converter does not isolate the piano from vocals, bass or drums. Everything pitched is transcribed together.
Tips for the best piano to MIDI results
- Use a solo piano recording. A digital piano's direct line output or a close microphone is ideal.
- Use less pedal and less reverb if you are recording specifically to transcribe.
- Start with a short section (30–60 seconds) to judge the result before converting a whole piece; the maximum is 3 minutes per file.
- If your digital piano or keyboard has MIDI output, record MIDI directly instead — it will be exact. Audio-to-MIDI is for when you only have the recording.
After conversion: making the MIDI usable
The file is written at a fixed 120 BPM, so the notes play at the right real-time positions but will not line up with bar lines. In most DAWs you can detect or tap the tempo and conform the MIDI, or simply set the project tempo to match the performance before quantizing. Then:
- Delete obviously spurious short notes and octave duplicates.
- Check note lengths where the pedal was down.
- Quantize gently — a rubato performance will be distorted by hard quantizing.
- For sheet music, import the cleaned MIDI into a notation program such as MuseScore and split it into two staves. The converter does not produce sheet music itself.
Recording is not piano-only? Switch to the general audio to MIDI converter (see also audio tools).
Piano to MIDI accuracy: what this converter got right in our 8-clip test
On eight piano clips with known notes, this converter found 118 of 125 notes (94%) with a median onset error of 4.6–9.3 ms per clip, but it also wrote 77 extra notes: single-line playing came out nearly clean, chords roughly doubled the note count, and three of the lowest keys (A0–E1) and four of the highest (C7–C8) were missed.
Method (2026-10-19): we wrote eight MIDI files with known pitches and start times, rendered them with FluidSynth 2.4.4 and the FluidR3_GM Acoustic Grand Piano, and ran each through this site's own note-detection worker (/vendor/transcribe.js, the Basic Pitch model, unchanged settings) in Chrome, fed 22,050 Hz mono audio exactly as the site prepares it. A note counts as correct when the pitch matches and it starts within ±50 ms of the true note; each true note can be matched once. Two full runs gave byte-identical output. The MIDI fixtures, scripts, raw output and scores are in results.json and the ground truth.
| Clip type | True notes | Found | Missed | Extra (harmonic) | Recall | Precision | Onset error ms, median / max |
|---|---|---|---|---|---|---|---|
| Slow melody (C major scale, quarter notes) | 15 | 15 | 0 | 3 (1) | 100% | 83% | 5.4 / 9.8 |
| Fast run (16th notes, 8 per second) | 15 | 15 | 0 | 3 (0) | 100% | 83% | 4.6 / 23.6 |
| Block triads (3 notes) | 24 | 24 | 0 | 21 (11) | 100% | 53% | 9.0 / 28.9 |
| Melody + bass line (two hands) | 24 | 24 | 0 | 12 (5) | 100% | 67% | 6.8 / 20.6 |
| Low register (A0–C3) | 10 | 7 | 3 | 17 (9) | 70% | 29% | 7.5 / 34.8 |
| High register (C6–C8) | 9 | 5 | 4 | 1 (0) | 56% | 83% | 7.5 / 11.6 |
| Repeated note (C4, 4 per second) | 12 | 12 | 0 | 1 (0) | 100% | 92% | 6.5 / 14.7 |
| Four-note seventh chords | 16 | 16 | 0 | 19 (3) | 100% | 46% | 9.3 / 19.1 |
What to expect from your own piano file: melodies, fast runs and repeated notes keep every note, with one to three stray notes to delete. Chords keep every note too, but expect two to five extra notes per chord (2.6 per triad, 4.8 per four-note chord here), some of them an octave or a twelfth above a real note; delete short notes and octave duplicates first. Notes at E1 and below, and from C7 up, may be missing (A0, C1, E1, C7, G7, A7 and C8 were missed; E7 was found) and need to be added by hand. Limits: these are clean synthesized clips with no pedal, room sound or other instruments, so a real recording will score lower; the test measures pitch and start time only, not note length or velocity.
Limits that apply to every file
- One file at a time, up to 50 MB and 3 minutes. Trim longer recordings into sections first.
- Mono analysis. The audio is mixed down to one channel at 22,050 Hz before note detection. If the left and right channels cancel each other out, see why a file that plays can be reported as silent.
- One instrument track, fixed tempo. All detected notes go on a single MIDI track. The file is written at 120 BPM; the converter does not detect the song's tempo, key or time signature, so bar lines in your DAW will not line up with the music until you adjust them.
- No pitch-bend data. Each detected note is written as a plain note-on/note-off with a velocity based on its loudness.
Today's captured length and size caps vs this page's 50 MB / 3 minute on-device limit
This page's 50 MB and 3 minute cap is the shortest free window in today's captured top 10, yet it is the only one that needs no upload and no account. Every other tool in the captured set either asks for a file upload, a sign-in, or both before it will return a MIDI file. The ratios below compare each captured cap against the 50 MB and 3 minute limit already stated on this page. They are arithmetic on published figures, not a ranking of audio quality. A longer window is only useful if you are willing to hand the recording to a server. This page keeps the audio in the browser, so the shorter window is the trade for staying local. The table lists only the caps that were readable on today's fetched bodies.
| Tool | Captured cap | Ratio vs this page | Upload or local |
|---|---|---|---|
| Eldoraudio (free) | 30 seconds | 180 / 30 = 6.0× longer here | Upload plus Google login |
| Eldoraudio (paid) | 30 minutes, 100 MB | 30 / 3 = 10.0× longer; 100 / 50 = 2.0× file cap | Upload |
| Jukeblocks | 5 minutes, 100 MB | 5 / 3 = 1.67×; 100 / 50 = 2.0× | Upload |
| Lacuna | 10 minutes, 30 MB | 10 / 3 = 3.33×; 50 / 30 = 1.67× | Upload, kept 30 days |
| AI-MIDI | 5 minutes, 50 MB | 5 / 3 = 1.67×; 50 MB equals 50 MB | Upload, sign-in |
| Ivory | 15 MB | 50 / 15 = 3.33× | Upload or YouTube link |
| PureMIDI | 25.0 MB | 50 / 25 = 2.0× | Upload, free-account preview |
| SoundToMIDI | 100 MB on-device | 100 / 50 = 2.0× | Stays on device |
Eldoraudio's free window is 30 seconds, so this page's 3 minute cap runs 180 / 30 = 6.0× longer for a free local run. Eldoraudio also requires an upload and a Google login, which this page does not. Its paid Flex and Pro tiers convert up to 30 minutes per file, which is 30 / 3 = 10.0× longer than this cap. That paid window still needs an upload, and its 100 MB ceiling is 100 / 50 = 2.0× this file cap. Jukeblocks allows 5 minutes and 100 MB, giving 5 / 3 = 1.67× on time and 100 / 50 = 2.0× on size, and it is an upload tool. Lacuna allows 10 minutes and 30 MB, which is 10 / 3 = 3.33× on time while this page's 50 MB is 50 / 30 = 1.67× Lacuna's file cap. Lacuna also keeps uploaded audio for 30 days, a retention detail this page does not have because nothing is uploaded. AI-MIDI allows 5 minutes and 50 MB, so 5 / 3 = 1.67× on time and an exact match on size, but it needs a sign-in and gives one full MIDI download each week. Ivory's 15 MB ceiling means this page's 50 MB is 50 / 15 = 3.33× the file cap, and Ivory also accepts a YouTube link and exports MusicXML, which this page does not. PureMIDI's 25.0 MB limit makes this page's 50 MB exactly 50 / 25 = 2.0× larger, and PureMIDI is an upload with a free-account preview. SoundToMIDI's 100 MB on-device cap is 100 / 50 = 2.0× this file cap, and it is the one other captured tool that keeps audio on the device. None of the top 10 put this 6.0× / 10.0× / 1.67× / 3.33× / 2.0× grid next to a labeled 50 MB / 3 minute no-upload cap. No price or unpaid remainder length is stated here because none was captured today.
Eldoraudio's "near-perfect" wording vs this already-published 118/125 (94%) run
Eldoraudio's captured page says "near-perfect accuracy" but publishes no recall, precision, or onset figure, so it cannot be compared to this page's 118 of 125 notes found. This page already reports that run from 2026-10-19 and rounds 118/125 = 94.4% to 94% recall. The same run found 77 extra notes, which is 77/125 = 0.62 extras per true note. The missed notes are 7/125 = 5.6% of the true notes. Median onset landed at 4.6–9.3 ms per clip across the eight clips. Those figures are cited here as already published and the eight-row table is not reprinted. A wording claim and a measured run are different kinds of evidence, and only one of them carries a number. The table sets the published result beside what each captured tool actually states.
| Source | What is stated | Numeric score on captured page |
|---|---|---|
| This page (run 2026-10-19) | 118 of 125 notes found, 94% recall, 77 extra notes, median onset 4.6–9.3 ms | Yes |
| Eldoraudio | "near-perfect accuracy" | None captured |
| MVSEP | Notes, timing and dynamics | None captured |
Eldoraudio's phrase "near-perfect accuracy" appears on the captured page without a recall, precision, or onset-ms figure beside it. MVSEP describes AI piano transcription of notes, timing and dynamics but also carries no numeric score on the fetched body. This page, by contrast, names 118 of 125 notes found and 77 extra notes on the same 125 true notes. The 94% recall figure is a rounded form of 118/125 = 94.4%, and the 5.6% miss rate is 7/125. The 77 extras work out to 77/125 = 0.62 extras per true note, which is a precision-side detail a bare accuracy claim never shows. Median onset of 4.6–9.3 ms per clip is the timing figure this page already publishes. None of today's top 10 place a "near-perfect" or score-less claim next to a named 118/125 and 77-extra result. This section does not claim Eldoraudio scored 94%, because no score was published on the captured page. It also does not claim Eldoraudio is worse, only that its wording cannot be checked against a number here. A reader comparing tools should treat an unquantified "near-perfect" as a marketing phrase until a recall or precision figure appears. The honest comparison is between a published run and a published claim, and only the run has digits attached. Anyone can rerun the same eight clips and check the 118/125 count for themselves. Until a tool publishes its own recall and extra-note count, the 94% and 0.62 figures here remain the only measured pair in this comparison.
Sustain pedal and pitch-bend claims vs what this converter actually writes
Lacuna's captured page promises sustain pedal in its .mid and Basic Pitch's demo advertises pitch bend detection, while this converter writes neither CC64 nor pitch-bend data. This page writes note-on and note-off events plus velocity, and nothing else. That limit is already stated on the page and is repeated here only to set it against the claims. Lacuna says its .mid includes velocity and sustain pedal, which is a broader event set than this converter produces. Basic Pitch's captured demo advertises pitch bend detection, and this page uses that open-source model but still writes no pitch-bend data. Ivory exports MusicXML, sheet music, and lead sheets from a YouTube link, none of which this page produces. The table lines up each claim with what this converter actually writes.
| Claim or feature | Captured source | What this converter writes |
|---|---|---|
| Sustain pedal in the .mid | Lacuna | No CC64 sustain |
| Pitch bend detection | Basic Pitch demo | No pitch-bend data |
| MusicXML / sheet music / lead sheets | Ivory | No sheet music export |
| Velocity | Lacuna, this page | Velocity written |
Lacuna's captured page says the .mid includes velocity and sustain pedal, so its output carries a controller event this converter does not write. This page writes velocities but does not export CC64 sustain, a limit already listed on the page. Basic Pitch's captured demo advertises pitch bend detection, and this page runs that open-source model yet still writes no pitch-bend data. The output here is note-on and note-off events plus velocity, with no continuous controller stream. Ivory exports MusicXML, sheet music, and lead sheets and accepts a YouTube link, while this page does not produce sheet music at all. The AIO note describes a physical digital piano or keyboard going over USB-MIDI or 5-pin into a DAW such as Ableton, Logic Pro, or FL Studio. That cable-and-DAW route is a different job from transcribing an audio recording, and it is not a missing tool on this page. A player who needs pedal or bend captured should record MIDI from the keyboard directly rather than expect it from an audio transcription. None of the top 10 put Lacuna's sustain-pedal claim and Basic Pitch's pitch-bend claim next to this converter's honest no-CC64, no-pitch-bend, no-sheet-music limits. No CC64 or pitch-bend test number is invented here because none was measured. The point is simply that a claim of pedal or bend output and a converter that writes only notes and velocity are not the same product. Reading the event list before choosing a tool saves a re-recording later. If pedal and bend matter to your arrangement, treat this page's note-and-velocity output as the starting point and add the controller data in your DAW.
Questions
Can it transcribe piano chords?
Yes, the model detects multiple simultaneous pitches, so chords produce several notes. Dense chords, pedal and reverb can cause missing or extra notes that need editing.
Will left and right hand be on separate tracks?
No. All notes are on one track. Split them by pitch in your DAW or notation editor.
Does it export sustain pedal?
No. Only notes with velocities are written. Add pedal (CC64) yourself if needed.
Can I convert piano from a full song?
You can, but the converter does not isolate the piano. Vocals, bass and other instruments are transcribed into the same track.