Transcribe audio into sheet music in your browser: the Basic Pitch neural network finds the notes in humming, singing, solo instruments, piano or a full song, then the tool tracks the beat through rubato, detects the meter (4/4, 3/4, 6/8, 12/8) and the downbeat, writes triplets and chord symbols, names the key, and engraves the result with tuplets, ties and a metronome mark; a song comes out as a piano arrangement, the tune over its bass line. A piano roll drawn on the tracked beats lets you remove or move notes; every setting re-computes instantly without re-running the model. Exports MusicXML (with tuplets, ties and harmony), MIDI, SVG, PNG, and PDF or print on Letter or A4 paper with whole lines of music on every page.
Initializing in your browser…
Turn audio into a waveform image in five styles, exported as PNG or real vector SVG. Reads every channel, shows true level in dBFS, and redraws correctly at any size.
Generate band-limited test tones, sweeps and noise from 20 Hz to 20 kHz. Up to eight voices, note-name entry, binaural beats, and WAV export rendered from the same oscillators you hear.
Detect tempo from onset autocorrelation, within 0.20 BPM across thirteen known-tempo fixtures. Draws the beats it found and plays them back as a click so you can check by ear.
A waltz recorded on piano: melody over an oom-pah-pah left hand in F major at 90 BPM.
Input
waltz.wav - Piano or chords - 34 s
Output
3/4 at 90 BPM, F major, grand staff, chord symbols F C Bb, 146 of 146 notes in the right bar and beat
The figures are measured on a recording synthesised from a written score, so every note, the meter, the downbeat and the chord under each bar were known before the tool ran. The build this replaces found the same tempo but barred the waltz in 4/4, placed 12 percent of the notes correctly and reached a note F1 of 0.49, mostly harmonic ghosts and bass notes eaten by an octave filter; now the accent pattern at the eighth level reads 3/4, the ghosts are filtered, and 16 of 16 bars carry the right chord. On a sung line with real rubato the same pass took correctly placed notes from 17 to 97 percent by tracking the beat instead of assuming one tempo.
Audio-to-sheet-music transcription has two halves. The first is hearing which pitches sound when: this tool runs Spotify's Basic Pitch convolutional network in your browser (a 900 KB model, downloaded once), which emits note-onset, note-active and pitch-contour probability maps at 86 frames per second for all 88 piano keys. The second half is everything that turns those note events into a readable page, and that is where this tool does most of its work: repairing note boundaries the network gets wrong, tracking the beat through a performance that speeds up and slows down, deciding whether the bar is 4/4, 3/4, 6/8 or 12/8, finding the downbeat, writing triplets where the notes fall in thirds, naming the key, estimating a chord symbol for each bar, and engraving the result with tuplets, ties, beams and a metronome mark. Every stage was measured against recordings synthesised from a written score, so the right answer was known before the network ran: a piano melody with repeated notes, a waltz over an oom-pah-pah left hand, a tune in 6/8 with a pickup, a line mixing triplets with straight eighths, a sung melody with vibrato and real rubato, block chords under a melody, a stereo mix with a lead vocal over a pad, an arpeggio, bass and a kick, and then three that carry what real records carry: a band in 4/4 with a sung phrase answered by an accordion over bass, kick, snare, hi-hat and a strummed chord on every eighth; a ballad in 12/8 with piano triplets, a ride and reverb on the voice; and a mono recording with no stereo image to isolate. On the ten, the note-level F1 is 0.90 to 1.00 on every one, the fraction of notes landing in the right bar and beat is 91 to 100 percent, the tempo and the meter are right on all ten, and the chord under every one of the 64 bars that have one is right. Synthesised fixtures are kinder than records, so the Full song setting is also measured on the first 90 seconds of 13 commercial songs (Balkan folk-pop, Adele, Ed Sheeran, Billy Joel, Metallica, the Animals and others) against independent piano transcriptions aligned to the recordings bar by bar, counting a written note as right when its pitch matches within a tenth of a second. That set chose how a song is arranged: the tune in the right hand, written an octave up when it sits below E4, and placed on the beat all at once; a bass line in the left hand, which the earlier lead sheet did not have; and chords decoded beat by beat. Together they took the melody's F1 from 0.12 to 0.22, the bass from nothing written to 0.31, and the chords from 0.41 to 0.51 (roots alone 0.46 to 0.58). Rules that did not raise those numbers here were left out. The same set shaped the Easy arrangement a song starts in: it leaves out the notes most likely to be wrong (a held note the network re-fired, a slide between notes), writes the tune on eighths and changes chord less often, for 39 percent fewer melody notes and a sixteenth on 8 percent of them instead of 32, at a melody F1 of 0.21. Slow songs gain from it and fast ornamented ones lose their runs, so Detailed is one click away. Every one of these choices comes with a control that lets you override it when a recording behaves differently.
Hum or sing it into the microphone and get a lead sheet you can edit in MuseScore or drop into a DAW as MIDI.
Upload a flute, violin, sax or synth lead and get its melody as notation for practice or arrangement.
Pick the chorus of a track and let Full song isolate the vocal and bass: the tune goes in the right hand, the bass line in the left and the chord symbols above. Switch the staff to Treble for a plain lead sheet.
Download a PDF laid out for Letter or A4 with whole lines of music on every page, or print straight from the tool; a larger staff size reads more easily from a music stand.
Keep every voice on two staves with chord symbols above, or switch to a single line for just the tune.
Transpose the score; the key signature follows so it stays readable.
Export MusicXML and open it in Finale, Dorico or Sibelius with tempo, meter, ties, triplets and chords intact.
Note extraction starts from Basic Pitch's own decoding and then repairs three faults that were measured on the fixtures. For a single line (a voice, a solo instrument, the lead of a song) the tune is not chosen among the polyphonic notes at all: it is decoded as one path through the network's note activations, where staying on a pitch is free, changing pitch costs a little more the further it leaps, and leaving or re-entering silence costs too, so the line that explains the most activation while moving the least is the tune. Measured on five 1980s and 1990s records with a voice or an accordion over a rhythm section, picking the loudest note at each instant gave three to five notes a second with ten-millisecond fragments an octave away; the path gives two to three coherent notes a second with none under 80 ms. The bass line is detected separately and its octave is pushed down in the activations before the path is found, so the tune does not drop onto the bass where the voice is quiet, and a note that resumes on the same pitch after a brief louder interruption (a chord stab under a held note) is kept as one note. Repeated pitches were being fused: a smoothing pass meant to heal a held note the network had dropped and re-fired also swallowed genuine repeats, so "C C G G" came back as "C G" and 18 of Twinkle's 42 notes were lost. The merge now asks the onset head and the loudness envelope whether the second fragment was struck. Legato repeats were never separated: a singer re-articulating the same pitch registers at 0.3 to 0.5 in the onset head, under the threshold that keeps vibrato from spawning notes, so a held pitch is now split where a clear onset peak follows a dip in the audio. And harmonics were transcribed as notes: on chords the network fired on the octave, twelfth and double octave of a struck note, 170 extra notes on 79 real ones, so a note that starts with a louder note a harmonic interval below it and lives inside its span is dropped as a partial. The pitch-contour head also gives each note its intonation in cents (calibrated against notes synthesised exactly in tune), which is shown on the piano roll and used to spell a note sung more than half a semitone sharp or flat as its neighbour.
Rhythm is a beat map rather than a tempo. A spectral-flux onset signal from the whole mix (not the isolated stem, which has lost the strums and hi-hats that carry the pulse) is fused with the note onsets, a tempo is estimated from their alignment on the sixteenth and eighth grids with the audio's own periodicity weighted by how clearly it repeats and a prior toward the tempos a page is written at, and a dynamic-programming beat tracker (after Ellis, 2007) finds beats that sit on onsets while keeping near that tempo, so a performance that leans from 84 to 96 BPM and back still lands on the right beats. When the audio itself keeps steady time (drums, strums) or the tracked beats barely vary, the score uses an evenly spaced grid instead; a control lets you force either. The meter is read from the accent pattern at the eighth-note level, which is the level at which a waltz and a jig differ: each candidate signature has a template of expected accents per eighth, accents come from note loudness, length and register, from onsets of the bass line, and from where the chord root changes, and the tracked pulse is tried as a quarter, an eighth and a dotted quarter. Which subdivision the off-beat notes fall on decides between a quarter pulse split in thirds and true compound time, and clean duple subdivisions argue for 4/4 with triplets over 6/8 with duplets. The downbeat is the bar shift that best matches the template, with a small bonus for the opening note, and notes before it form a pickup bar.
Quantization measures each note against the beats it fell between, in twelfths of a quarter note so that sixteenths (3) and triplet eighths (4) are both whole numbers. Each beat picks its own grid: the duple grid, or a triplet grid when two onsets in the beat sit clearly off the duple grid and fit thirds. On the Auto setting a single line is placed all at once rather than note by note: every onset may move to a sixteenth within a sixteenth and a half of where it was heard, a sixteenth off the eighth grid costs extra, and no two notes may share a slot, so lilted eighths are written as eighths, real sixteenth runs stay sixteenths, and two quick notes are no longer rounded onto one eighth where one of them was lost. Note lengths are held by default for a voice or a song: a note lasts until the next one unless the gap is longer than three beats, and a phrase that stops ends on the next beat, which cut the rests in a song's melody from 0.9 to 0.2 a bar. The engraving grammar then splits and ties values where a reader needs to see the beat: a half note may not start on beat two of 4/4, an eighth may not straddle a beat, halves and wholes do not exist in 6/8, and a triplet beat is written in tuplet values under a bracket or a beam. Chord symbols are decoded one beat at a time: each beat's pitch-class profile from every note heard (including the bass line, which is detected regardless of the pitch range you set for the staff) is compared with the 24 major and minor triads, the bass note sounding on a chord's root counts for it, and a Viterbi pass charges a cost for every change of chord, a fifth as much on a bar line, so a passing note does not become a symbol while a change on the bar always can. Sevenths are written as their triads, which matched the reference transcriptions better than guessing the seventh. In Melody + bass, the score's layout for a song, the left hand is one bass line: of the notes the network hears in the bass register the loudest of each cluster of onsets, kept near the register the line has been in, folded into the octave from E-flat 2, on the eighth-note grid, with a repeat that is still sounding held rather than re-struck. Rendering is VexFlow with tuplets, beams, cross-line ties, chord symbols and a metronome mark (a dotted quarter in compound meters). MusicXML 3.1 carries all of it (time-modification and tuplet notations, tied notes, harmony elements, key, meter and tempo) into MuseScore, Finale, Sibelius or Dorico; MIDI is a format-0 file at 480 ticks per quarter with key and time signature meta events and the chord symbols as text; SVG and PNG come from the on-screen rendering; PDF and print engrave the score again at the width of the chosen paper (Letter or A4, at three staff sizes), measure every system as drawn, and lay out whole systems per page, spread to fill every page but the last, with page numbers, so no staff is cut at a page edge. Because the network's output is cached, every detection, rhythm and notation control re-computes in tens of milliseconds; only a different section of the file, or switching centre isolation, runs the network again.
Basic Pitch is the one lightweight, instrument-agnostic, polyphonic transcription model with weights that run in a browser; the alternatives (Onsets and Frames, MT3, the piano-only transformers) are piano-specific, tens of megabytes, or both. Measured on fixtures with a written answer, the network found every note of a piano melody at the raw level; the errors were in what came after it, which is why this tool's work is in boundary repair, beat tracking, meter and engraving rather than in swapping the network.
Yes, with the Full song setting, which isolates the centre of the stereo image (lead vocal, bass, kick) before the network hears it, decodes the lead as one coherent line through the network's activations, and puts chord symbols above from everything heard. On the three band fixtures it was measured on (a sung phrase answered by an accordion over drums and a strummed chord, a 12/8 ballad with reverb on the voice, a mono record with nothing to isolate) 90 to 97 percent of the notes are right and 91 to 98 percent land in the right bar and beat. Real records are harder still: a doubled vocal is two lines to the network, a harmony singer may take the tune for a phrase, and a repeated note sung without a consonant between is one note to the network and to a listener. Press Play together to hear the transcription over the recording, then fix what is left on the piano roll.
A groove whose eighths group in threes can be read as a dotted-quarter pulse (6/8 or 12/8 at 85) or as a quarter pulse of the same eighths (4/4 at 127), and some recordings carry both periodicities at nearly equal strength: a 3+3+2 pattern in 4/4 and a 12/8 shuffle put the same peaks in the audio. The tool reads the accent pattern, the off-beat subdivisions and the chord rhythm and reports how clear the decision was; when it goes the other way, choose the meter in the Rhythm tab and the tempo re-reads itself (a dotted quarter at 85 is the same pulse as a quarter at 127).
It quantizes each note against the beats the tracker actually heard, so a performance that speeds up and slows down still notates cleanly: on a sung line whose tempo leaned from 84 to 96 and back, 17 percent of notes were in the right place on a rigid grid and 97 percent following the beats. Strict grid spaces the beats evenly at one tempo, which is right when the accompaniment keeps time. Auto picks by measuring how steady the tracked beats and the audio's own pulse are.
The network's pitch-contour head resolves pitch to a third of a semitone. The tool measures how far each note sat from its written semitone, in cents, after calibrating against notes synthesised exactly in tune. A note sung more than half a semitone away from the written pitch is spelled as its neighbour with the remaining offset.
The 900 KB model has to be downloaded and compiled for TensorFlow.js the first time; after that it is cached. Inference itself runs at several times real time on the graphics processor. Everything after inference is cached, including the network's own note decoding per sensitivity and pitch range, so changing a setting re-computes the page in a few tens of milliseconds (measured: 16 ms for a 30 second region) and never runs the network again; only a new section of the file, or switching centre isolation, does.
Key detection correlates the duration-weighted pitch-class profile against the 24 major and minor key profiles, with a small bonus for the pitch the excerpt ends on. It needs twenty or more notes of tonal material; short, modal or chromatic excerpts throw it off, and a relative major and minor share a signature and often score within a hair of each other. The Notation tab lets you set it.
Yes. MusicXML preserves notes, durations, ties, triplets, key, meter, tempo, chord symbols and (for a grand staff) the staff of each note, so MuseScore (free), Finale, Sibelius and Dorico open it for editing. The piano roll also lets you remove or move notes before you export.
Yes. Under the export buttons, choose Letter (8.5 × 11 in) or A4 (210 × 297 mm); the choice is remembered, and the first time it follows your browser's region. PDF and Print lay the score out for that paper: it is engraved at the width of the page, every line of music is measured as drawn, and a line that does not fit on a page goes to the next one whole, so no staff is cut in half at the bottom. The staff size (compact, standard or large) changes how many bars fit on a line and how many lines on a page.
A song starts in the Easy arrangement, which writes fewer notes so the page can be played. It leaves out two kinds of note that were mostly wrong when checked against independent transcriptions of 13 songs: a repeat of the pitch just sung that has no clear new attack (usually one held note the network heard twice), and a change of pitch with almost no attack (a slide into a note). It also writes the tune on eighths and changes chord less often. Fast runs come out simplified. Switch Arrangement to Detailed in the Notation tab to see every note, with sixteenths where the timing says so.
In the Melody + bass layout the tune is the right hand of a piano part, and a low voice sung around middle C would sit on top of the left hand with ledger lines below the treble staff. When the middle note of the tune is below E4 it is written an octave up, which is also how treble-clef parts for low voices are written. Measured against independent piano transcriptions of 13 songs, it raised the melody's F1 from 0.13 to 0.22. Set Tune to As sung in the Notation tab to keep the sung octave.
No. The model, the note extraction, beat tracking, meter and key estimation, chord symbols and the rendering all run in your browser. The audio never leaves your device.
Audio is decoded and processed locally with the Web Audio API. Your files are never uploaded to a server.