Music Similarity Checker
Music similarity checkers like this one compare two melodies or rhythmic patterns and return a score or percentage. But…
Does a high score mean infringement? Maybe.
Does a low score mean no infringement? Maybe.
So Musicologize built the most comprehensive toolbox we know of, and it at least tries to run a valid experiment. Put in two melodies and 29 published similarity algorithms each give you a score.
Every one of those algorithms is legitimate. Each does exactly what it was designed to do, whether that’s counting the notes you’d have to change to turn one melody into the other, lining the two up note against note, or comparing their rhythms. As far as each one goes, it’s right. But they measure different things, so they disagree, sometimes wildly. The same two melodies can score close to nothing on one and close to identical on another. None of them was designed to measure copying.
A score on its own also means very little. So every score is checked against hundreds of pairs of unrelated songs, because “80% similar” means nothing until you know how similar unrelated songs usually come out. That’s the experiment. Read the field guide
1Enter your two melodies
Everything runs in your browser. Nothing you enter is sent anywhere. Up to 64 notes per melody.
Why no checker can measure infringement
If you want to know whether a song infringes, a similarity checker is a waste of your time. Mine included. Don’t take my word for it. Load the examples above and watch.
A percentage can’t tell a stolen hook from a scale.
Every algorithm here compares notes, intervals or rhythms. None of them knows which notes matter. A note that matches is a note that matches, whether it’s part of a hook nobody else wrote or a run up the scale that’s in thousands of songs.
The algorithms don’t agree with each other.
Put the same two melodies through all 29 and you get everything from almost nothing to almost identical. Twinkle against Ode to Joy scores anywhere from 0% to 99%. Each algorithm is right about the thing it measures. None of them measures copying.
Familiar melodies come out “similar.”
Simple melodies that step up and down a major scale sound like each other. Against popular songs, 15 of the 28 algorithms that run on Twinkle and Ode to Joy call that pair more alike than 95 out of 100 unrelated pairs. That proves nothing, except that simple tunes are simple.
A “plagiarism risk” percentage is a similarity score with the word risk printed on it.
Checking your upload against millions of recordings will find songs that sound like yours. It won’t set aside what everybody in your genre does. It won’t ask what came first. It won’t ask whether the part that matches is something anyone can own. Those are the questions an infringement claim actually turns on.
Whether a similarity matters comes down to what’s shared, whether anyone can own it, and what came before both songs. That takes a musicologist, working note by note.
Field guide
These are serious tools with real histories, and each one does what it was built to do. Eight of the 29 were designed for music. The rest were borrowed from other fields: protein and DNA research, speech recognition, ecology, psychology, statistics, information theory and computer vision. Not one was designed to decide whether one song copied another. Using them well means knowing what each one asks, and what it can’t see.
Edit distance, four ways
One algorithm, used four ways. Edit distance counts how many notes you’d have to add, delete or change to turn one melody into the other. The four versions compare note names, intervals, ups and downs, and note lengths. Many similarity checkers stop here.
Levenshtein: note names
Coding theoryNamed for Vladimir Levenshtein, who defined the distance in a 1965 paper on error-correcting codes.
How many single-note edits turn one pitch sequence into the other?
What it’s good for
It’s fast and easy to follow. It spots near-identical melodies even with a few wrong notes, which makes it good for finding a tune from a sloppy transcription.
Where it goes wrong
A wrong note costs the same whether it’s a half step off or a fifth off. Rhythm is ignored completely.
Levenshtein: intervals
Coding theoryLevenshtein, 1965.
How many edits turn one sequence of intervals into the other?
What it’s good for
Because it compares intervals, key doesn’t matter. The same tune in another key scores as identical.
Where it goes wrong
Change one note and two intervals change, so a single wrong note counts double. Rhythm is ignored.
Levenshtein: contour
Coding theoryLevenshtein, 1965.
How many edits turn one up, down and repeat pattern into the other?
What it’s good for
The ups and downs are what most listeners remember. It forgives wrong notes as long as the melody moves in the same directions.
Where it goes wrong
Up, down and repeat is all it sees. Thousands of melodies share the same ups and downs.
Levenshtein: note lengths
Coding theoryLevenshtein, 1965.
How many edits turn one sequence of note lengths into the other?
What it’s good for
A quick first look at whether two melodies share a rhythm, ignoring pitch.
Where it goes wrong
It sees only note lengths. Any two tunes in steady quarter notes look identical to it.
Parsons code: Levenshtein
Built for musicNamed for Denys Parsons, whose 1975 Directory of Tunes and Musical Themes indexed melodies by contour alone.
Do the two melodies go up, down and repeat in the same order?
What it’s good for
Parsons code is what a person can write down after hearing a tune once. It has long served as a way to look up a melody from memory.
Where it goes wrong
It throws away how far the melody moves and how long each note lasts. Very different tunes can have the same contour.
Alignment
Alignment lines two melodies up note against note and scores every match, mismatch and gap. Most of these methods were invented to compare proteins and DNA. They work on music because a melody, like a strand of DNA, is a sequence in a fixed order.
Needleman–Wunsch alignment: note names
Molecular biologySaul Needleman and Christian Wunsch published it in 1970 to compare the amino-acid sequences of proteins.
What is the best end-to-end alignment of the two melodies, note against note?
What it’s good for
It is built to compare two whole sequences of similar length. It shows exactly which notes correspond and where the gaps fall.
Where it goes wrong
It insists on matching the whole melody end to end, so a short shared phrase inside two otherwise different songs gets buried.
Gotoh alignment: note names
Molecular biologyO. Gotoh’s 1982 refinement for matching biological sequences.
The same end-to-end alignment, where one long gap costs less than many short ones.
What it’s good for
This is closer to how melodies are varied in practice. A phrase is often inserted or dropped whole rather than note by note.
Where it goes wrong
It also matches whole melodies end to end, so a short shared phrase gets buried the same way.
Smith–Waterman local alignment: note names
Molecular biologyTemple Smith and Michael Waterman, 1981, for finding common subsequences in molecules.
Where is the single best-matching passage inside the two melodies?
What it’s good for
It is the right tool when a short shared phrase sits inside two otherwise different songs, which is the usual shape of a real dispute.
Where it goes wrong
It finds only the single best-matching passage. Two melodies that share three short motifs score no higher than two that share one.
Longest common subsequence: note names
Computer scienceoriginator, 1975.
What is the longest run of notes both melodies share in the same order, with gaps allowed?
What it’s good for
It forgives added passing notes and ornaments, because extra notes in between do not break the match.
Where it goes wrong
Because gaps cost nothing, it can string together notes scattered all over both melodies. Longer melodies find more matches by chance.
Longest common subsequence: intervals
Computer scienceoriginator, 1975.
The same longest shared run, measured on intervals instead of pitches.
What it’s good for
It forgives added notes, and the key doesn’t matter.
Where it goes wrong
The same weakness: free gaps let it collect intervals that line up by chance.
Dynamic time warping: pitch
Speech recognitionHiroaki Sakoe and Seibi Chiba, 1978, for spoken word recognition. T. K. Vintsyuk applied dynamic programming to speech ten years earlier.
How closely do the two pitch lines match if each may be stretched or squeezed in time?
What it’s good for
It was built for speech, where the same word is said fast or slow. It handles sung and performed input with uneven tempo gracefully.
Where it goes wrong
It stretches time to make the melodies line up, which means it throws rhythm away. The same pitches in completely different rhythms can score as a perfect match.
Mongeau–Sankoff weighted edit distance
Built for musicMarcel Mongeau and David Sankoff, 1990.
What is the cheapest way to turn one melody into the other, counting pitch and duration and letting one note split into two?
What it’s good for
It was one of the first similarity measures designed for music. It knows that a half note played as two quarter notes is a small change, not two edits.
Where it goes wrong
The cost of each kind of change was set by hand, by judgment, not measured. Change the costs and the answer changes.
Primary source Partly unverified
Shared fragments
These cut each melody into short runs of intervals and count how many the two have in common, wherever they turn up. They started in ecology, as ways to compare which species live in two places.
Interval n-gram Jaccard
EcologyPaul Jaccard, 1901, comparing the flora of regions of the Alps and the Jura.
Of all the short interval patterns in either melody, what share appear in both?
What it’s good for
It finds shared motifs wherever they occur and in any order. That makes it good at spotting a figure that has been moved to a new place.
Where it goes wrong
Order and repetition don’t count, and a longer melody drags the score down.
Interval n-gram Dice
EcologyLee R. Dice, 1945, measuring how often species occur together. Thorvald Sørensen reached the same formula independently in 1948.
The same overlap of patterns, weighted toward what the two share.
What it’s good for
It ranks pairs in exactly the same order as Jaccard, on a scale some people find easier to read.
Where it goes wrong
It has Jaccard’s blind spots: order and repetition don’t count, and melody length skews the result.
Interval n-gram Tversky
PsychologyAmos Tversky, 1977, from a psychological account of how people judge resemblance.
How much of one melody’s patterns are found in the other, counted in one direction?
What it’s good for
It can be made one-directional, to ask whether a short hook is contained in a longer song without charging the song for everything else it does. Here its two weights are set equal, which makes it match Dice. The setting is shown so the choice is visible.
Where it goes wrong
Someone has to decide which melody counts as the original, and that choice sets the answer.
Distributions and habits
These describe a melody by its habits: which notes it uses, which intervals, which move usually follows which. They compare habits, not tunes, and they need longer passages to say much.
Pitch-class histogram cosine
Information retrievalSalton et al., 1975.
Do the two melodies use the same notes of the scale in the same proportions?
What it’s good for
It gives a quick picture of which notes a melody leans on. It’s useful for comparing key and mode and for sorting music by style.
Where it goes wrong
Order doesn’t matter at all. Shuffle a melody’s notes and it still scores 100% against the original.
Interval histogram cosine
Information retrievalSalton et al., 1975.
Do the two melodies favor the same steps and leaps?
What it’s good for
It captures a melody’s characteristic intervals, such as a habit of leaping a sixth, independent of key.
Where it goes wrong
Order doesn’t matter. Shuffle the intervals and the score doesn’t change.
Kullback–Leibler divergence: note-to-note habits
StatisticsSolomon Kullback and Richard Leibler, 1951.
How surprised would a model of one melody’s note-to-note habits be by the other?
What it’s good for
It compares how two melodies move rather than which notes they use. It is one way to compare the habits of two repertoires.
Where it goes wrong
On short melodies it shoots to infinity the moment one melody makes a move the other never does, so the answer depends on a smoothing setting.
Jensen–Shannon divergence: note-to-note habits
Information theoryJianhua Lin, 1991. Sibson described the same quantity in 1969.
The same comparison of note-to-note habits, with a fixed range and the same answer in both directions.
What it’s good for
It fixes the Kullback–Leibler habit of shooting to infinity, so short melodies still get a sensible number.
Where it goes wrong
It only looks one note ahead, so phrases and structure are invisible, and a few dozen notes are too few to measure habits reliably.
Normalized compression distance
Information theoryLi et al., 2004.
How much does knowing one melody help a compressor shrink the other?
What it’s good for
It makes no musical assumptions at all, which is its appeal. On whole pieces it has been used to group music by genre and by composer.
Where it goes wrong
The answer depends on which compression program you use. There’s no fixed scale.
Shape and geometry
These draw the notes as shapes on a grid of pitch and time and compare the shapes. Several were designed for music from the start, and they handle a change of key as part of the geometry.
Earth Mover’s Distance
Computer visionYossi Rubner, Carlo Tomasi and Leonidas Guibas, 1998, for searching image databases. Adapted to melody by Rainer Typke and colleagues in 2003.
How much work does it take to move one melody’s notes onto the other’s, like moving piles of earth?
What it’s good for
It allows partial matches and small slips in pitch or timing without falling apart. It was one of the systems tested in the field’s shared benchmark.
Where it goes wrong
When one melody has more notes, or longer ones, than the other, it stops behaving like a true distance.
Proportional Transportation Distance
Built for musicRainer Typke and colleagues, 2003.
The same earth-moving comparison, with both melodies given equal total weight.
What it’s good for
Giving both melodies equal weight makes it a true distance, so it can search a large collection quickly.
Where it goes wrong
It lines up the key by average pitch rather than by the best fit, and grace notes need handling the method doesn’t spell out.
P3: longest common time
Built for musicEsko Ukkonen, Kjell Lemström and Veli Mäkinen, 2003.
If one melody is slid in time and pitch, how much of it lies exactly on top of the other?
What it’s good for
It copes with a change of key or a shift in time automatically, and it was designed to find a melody hidden inside a fuller arrangement.
Where it goes wrong
Only exact matches in pitch and time count. A note a half step off earns nothing at all.
Ó Maidín area difference
Built for musicD. S. Ó Maidín, 1998. Given its formal geometric treatment by Aloupis and colleagues in 2006.
If both melodies are drawn as pitch lines over time, how much area lies between them?
What it’s good for
It compares shapes directly, with no rounding of rhythms and no guessing at keys. That makes it a useful check on how much the other algorithms depend on those steps.
Where it goes wrong
It works in continuous pitch and time, so it knows nothing about meter or where the beat falls.
Rhythm
These ignore pitch. They look at where the notes start and how long they last. Several were developed to trace family resemblances among traditional rhythms, like the ones in flamenco.
Hamming distance: rhythm grid
Rhythm studiesHamming, 2004.
On a shared grid of pulses, how many positions differ between the two rhythms?
What it’s good for
It’s the simplest rhythm comparison there is: put both on the same grid and count the mismatches.
Where it goes wrong
A note one pulse late counts the same as a note a whole bar away.
Swap distance
Built for musicDíaz-Báñez, Farigu, Gómez, Rappaport and Toussaint, 2004, in a study of flamenco rhythms.
How many small nudges of single notes turn one rhythm into the other?
What it’s good for
It treats a note pushed a little early or late as a small change, which suits rhythms that differ by a few nudged notes.
Where it goes wrong
It only knows where notes start. It can’t tell which notes are accented, and it needs everything on a fixed grid.
Chronotonic distance
PhoneticsKjell Gustafson, 1987, at Oxford’s Phonetics Laboratory, as a way to picture the rhythm of speech.
Drawn as a chain of squares sized by duration, how different do the two rhythms look?
What it’s good for
It ranked first among rhythm measures in a 2004 comparison of how well each one grouped related rhythms.
Where it goes wrong
Long notes dominate the score, because each note’s weight grows with the square of its length.
Interval-ratio distance
Built for musicE. J. Coyle and I. Shmulevich, 1998.
Do the durations change from note to note by the same ratios?
What it’s good for
Tempo doesn’t matter, because a rhythm played twice as fast keeps the same ratios.
Where it goes wrong
It only works when both rhythms have the same number of notes, so it often can’t run at all.
Note-length histogram correlation
Built for musicTzanetakis & Cook, 2002.
Do the two rhythms use the same note lengths in the same proportions?
What it’s good for
It’s a quick read of which note lengths a melody uses, a cousin of the tools used to sort recordings by genre.
Where it goes wrong
It counts note lengths and ignores their order. Shuffle the rhythm and the score doesn’t move.
Frequently Asked Questions
Does this tool tell me whether my song infringes on another song?
No! And we’re sorry, but no tool can do that! Copyright infringement is not so simple; it is not merely a math problem; it is not just a percentage or a similarity score. Copyright infringement is a legal conclusion that depends on whether protectable expression was copied. A similarity percentage that comes out of a tool like this only one measures what I’d call “gross similarity” whereas a similarity percentage that comes out of a tool like this one is a measure of what I’d call “gross similarities” from a mechanical look at notes that can’t distinguish meaningful similarities from meaningless ones, original expression from blue-chip note similarities from scenes à faire. I could go on forever. If there’s a shortcut, I’d have found it. The tool produces smoke. Confirming or dismissing an actual fire is the domain of a forensic musicologist, one with training, judgment, and repertoire knowledge.
What does the similarity percentage actually mean?
Think of it as calculating gross similarity through a series of mechanical filters. The melody comparison transposes both inputs to a common key (that’s actually something I do too), then it effectively writes them on vellum paper, holds them up to the light (that’s what I do sometimes too), and makes some comparisons. It may measure the edit distance, which is a fancy way of asking how many notes you’d have to add, remove, or swap to turn one melody into the other. Fewer edits, closer match. The interval comparison looks at melody differently, measuring the distances between consecutive notes. The contour comparison simplifies the melodic arc to a zoomed-out view of whether the pitches go up, down, or stay the same. And the rhythm comparison quantizes tapped-out notes (snaps them to a fixed grid) and compares the rhythmic values by standard subdivisions. These are all isolated measurements of a single property. It’s data. Data is meaningful, but needs to be properly applied. Music is rarely one property at a time. As Albert Einstein might’ve put it, forensic musicology should be as simple as possible, but no simpler. Notes get their meaning from context, how they relate to each other, and to the harmony, and to the flow of time. You cannot find the musical truth with a tool like this. It would be scammy to suggest otherwise.
If the percentage is high, should I be worried?
Not necessarily. Not from this alone, certainly. I called this kind of thing “gross similarity,” but what you really want is something closer to net similarity, and a forensic musicology analysis would introduce the metaphorical net that filters away what’s less relevant. A high score from a tool that doesn’t consider the originality of the underlying material is a nothing-burger. I’ll try to come up with an illustration on the spot. (Thinks a minute.) Asked myself, “where else can I find the seven-note beginning of “Twinkle, Twinkle, Little Star?” First song that came to mind — had to look up whose it was — Tom Hanks in Bachelor Party, “Why Do Good Girls Like Bad Boys?” If you ran those seven notes through this little “Music Similarity Checker” you’d get a 100% match. concluded that the 100% match across those seven consecutive notes means they and Tom Hanks stole “Twinkle Twinkle?” Well, maybe, but they’d then remember Twinkle is really, really old and whatever. I’m getting off track here. The. Point. Is. people, that the 100% similarity there is pretty meaningless, whereas, conversely, a moderate sounding score on highly original material could still indicate copying. A percentage like this one tells you that the sequences of pitches line up. (It might also ideally have told you that so do the rhythms and the harmonic context, which also starts to sound like a lot, if you’re a “Music Similarity Checker,” but it does not tell you whether those identical notes are diamonds of forensic musicology or lumps of coal. Determining whether a similarity is significant requires expertise and rigor, forensic analysis that no mechanical tool provides.
How is this different from other music copyright checkers?
Fundamentally, it isn’t. When someone builds something that works, I’ll want to use it. In the meantime, any tool that uses the aforementioned edit distance or similar algorithms to compare pitch sequences or rhythmic patterns is performing the same basic gross sort of operation, whether it accepts audio input or uses a virtual piano, for now, makes no difference. The underlying math and its limitations don’t change: a mechanical comparison cannot determine whether copying occurred, or whether the shared elements are protectable, so they can’t know whether the similarity is legally meaningful. The Musicologize Music Similarity Checker is transparent about those limitations. That’s the point of it.
Why can’t a tool do what a forensic musicologist does?
Why can’t a radar gun on a pitcher do what an MLB baseball scout does? It’s amazing enough that it knows a pitcher throws mid-90’s, but it certainly doesn’t know if they can get hitters out. A forensic musicologist evaluates music the way it actually functions, and not merely as isolated sequences of pitch and rhythm. Musical elements derive their meaning from harmonic context, structural placement, interaction with other elements, and even relationship to the broader musical repertoire. I’m gradually running out of ways to illustrate it. A tool treats a three-note figure in a vacuum. A musicologist knows those notes can mean something entirely different at the beginning of a verse over a stable tonic chord than they do introducing a bridge over a pivot chord that modulates to a new key. An algorithm cannot weigh a melodic phrase against other works within and without the repertoire to determine whether it is original expression or a common building block that belongs to no one. These are judgments that require musical training, analytical experience, and knowledge of the case law. They are not reducible to a formula.
Can I use these results in a legal proceeding?
No, and doing so would likely be counterproductive. A similarity percentage from an online tool has no evidentiary value. Courts rely on a forensic musicology expert report, a courtroom-calibrated document that examines the works through exhibits built from forensic notation, parallel transcriptions, and prior art research. It helps to have the heart of a teacher — to build the explanations and illustrations that actually land with a judge or jury who don’t read music. Submitting a mechanical percentage without musicological context will not adeqately support your position. It is more likely to undermine it by focusing on what I’ve termed “gross similarity” when the law knows to ask about net similarity. If you need analysis that holds up under legal scrutiny, you need an expert, not a calculator.
Why offer this tool if it can’t answer the real question?
Because the question people bring to tools like this is one worth taking seriously, even if no algorithm can answer it. If your ears are telling you two songs sound alike and you want to know whether there is fire behind that smoke, a comparison like this can confirm that the similarity has a measurable basis in the melodic or rhythmic content. That is a useful first step. The problem arises when a percentage is treated as a conclusion rather than a starting point. This tool is a starting point. The conclusion requires a conversation with someone who does this work for a living.
When a contract, a policy, or a platform requires a clearance, a written, signed opinion from a qualified forensic musicologist is what satisfies it.
If you’re hearing something that concerns you, whether you’re worried someone copied you, worried you might have copied someone, or trying to clear a track before you release it, the next step is a conversation, not another tool. (212) 217-9512.
(212) 217-9512 · Schedule a call · The initial call is free.
If you want to check a song yourself before calling anyone, start here: How Do I Check if My Song Sounds Too Much Like Another Song?