String Distance Calculator

Type two strings and see how far apart they are under several edit-distance metrics at once. Levenshtein counts insertions, deletions and substitutions — the classic spell-check measure, where kitten and sitting are three edits apart. Damerau-Levenshtein additionally treats a swap of two adjacent characters as a single edit, which matches how people actually mistype: ac and ca are two edits under Levenshtein and one under Damerau. Hamming counts position-by-position mismatches and is only defined when the strings are the same length, which makes it the right tool for fixed-width codes and the wrong one for free text. The similarity percentage is (1 − levenshtein / longer length) × 100, which is the number you would threshold on when deduplicating records.

The comparison runs over UTF-16 code units, and that has consequences worth knowing before you trust a distance. An emoji outside the basic plane is two code units, so deleting one counts as two edits rather than one. More awkwardly, two strings that look identical on screen can have a non-zero distance: café written with a precomposed é is four units, and the same word with e followed by a combining acute accent is five, giving a distance of 1 between two things no reader would call different. Normalise both sides — String.prototype.normalize('NFC') — before comparing anything that came from different sources, because text pasted from a Mac, a database and a web form will not agree on this by itself.

Choosing the right metric matters more than tuning the threshold. Edit distance treats every character as equally important, so Smith and Smyth score the same as Smith and Smith., though one is a different name and the other is punctuation. For matching personal names, Jaro-Winkler weights a shared prefix more heavily and does better; for matching sentences, token-based measures beat character-based ones outright. And raw distance is not comparable across lengths — three edits in a five-character string is a different event from three edits in a hundred-character one, which is exactly what the percentage is there to correct for.