We are building a new computer system from the ground up, made to be usable by everyone. Every thing in it, whether a person, a shared space or a piece of work, is named by a long number that nobody could ever say aloud. So each one also gets a short spoken name, a few simple syllables you could read down a phone line to someone in another country. That only works if the syllables sound different to everyone, whatever language they grew up with. This test is how we find out which ones do.
Each section below stands on its own. Open the ones you care about.
In the system we are building, every thing has one long, exact name: a number worked out from the thing's own contents, far too long for anyone to say. So each thing also gets a koan, a short way to say that name out loud. We take the start of the number, turn each piece of it into one of 64 simple syllables, and say those. A couple of syllables are enough to name a person. Four name a shared space. Up to eight name one single thing. The length tells you what kind of thing you are hearing.
A koan is only a way of saying the number. Two people who pronounce it differently, or use different sets of syllables, still mean the same number. Saying it differently never changes the thing itself. That is what lets us change the syllables later, if testing shows we chose badly, without disturbing anything that has already been named.
There are two candidate sets of syllables. set A uses eight consonants (k t p m n s l w), the four vowels a e i u, and an optional “ng” ending. set B uses the same consonants, the vowels a o i u, and an optional “n” ending. Eight consonants times four vowels times two endings is 64 either way. The two sets differ in exactly the places we are not sure people can hear: e against i, o against u, a against o, i against u, and n against ng.
The first version of the alphabet had both r and l. For several billion people, among them speakers of Japanese, Korean and many languages of China, r and l are one sound, so half the alphabet would have collapsed into the other half. We learned that from a paper study, which scored written spellings read out by 55 computer voices and never used a single human ear. The person who started this knew it from experience.
Some pairs of sounds are hard to tell apart in some languages and easy in others. Spanish speakers often hear v and b as one sound, and much of Latin America hears s and z as one. In Japanese, Finnish and Tamil the length of a vowel changes the word, so a stretched vowel might or might not survive a bad phone line. e and i, or o and u, blur in plenty of accents. Nobody can know which pairs matter from their own language alone, so we are asking people from many.
The aim is simple: pick the 64 syllables that the most people, in the most languages, on the worst phone line, can still tell apart.
An earlier version of this test (one voice, English speakers only, 10 people, 330 answers) found that r and l were never confused: 0 wrong out of 30. That is exactly what you would expect from English listeners, and it tells us nothing about the people the test was built for. e against i was wrong 7 times out of 30, p against k 8 out of 30, and v against b 8 out of 30. Those are all small counts and could be chance.
Live numbers are on the results page.
You hear one syllable, and two spellings appear. You pick the one you heard. Each pair of sounds comes up twice, in different syllables, and the two buttons swap sides at random, so always pressing the left one cannot look like good hearing. There are 28 questions.
The sounds are spoken by the voices already built into your browser, and the test takes turns between all of them. We record the voice's name with each answer, because a computer voice reads a spelling its own way: a mistake is partly your ear and partly the voice. Many different voices, heard by many different people, let us tell the two apart. So treat your own score as a hint, not a verdict. Nothing here is a judgement of you: you are the measuring instrument, not the thing being measured.
At the end you see your score for each pair next to everyone's, and next to people who share your first language once there are enough of them. You also get a koan of your own: a random one, made on the spot in your browser and shown in both sets of syllables. It is not saved or sent anywhere. The syllables and spellings are never translated, because they are the test; only the words around them are.
Kept for each answer: the syllable played, the other choice, the one you picked, which pair of sounds was being tested, the voice (its name and language), the time, the language this page was shown in, and a random code made in your browser that ties one person's answers together. Kept once per visit, only if you typed it: your first language(s), other languages, and country or region, exactly as you wrote them. Your name is a separate choice at the end.
Not kept by the test: your IP address, your device or browser details, any cookie, any account. We do not look at where the request came from. The company that hosts the page, Cloudflare, keeps its own standard logs, which we do not read and cannot control.
How it is used: to decide which syllables go into the spoken alphabet, and to publish what we find. It is all open. The tables and a spreadsheet of every answer (CSV) are public. The spreadsheet holds the language and region text you typed, the voice, the answers, and the random code for your visit, and never a name. Please do not type anything into those boxes that you would not want published.
At the end you choose. Stay anonymous is the default: you are counted, never named. Or credit me as a name or handle of your choosing, which then appears on the credits page. The name is stored separately from your answers and is not in the public spreadsheet. We do not check names. Ones with links or code in them are refused, and we will remove any that is abusive.
This is a small part of a bigger piece of work, and everyone who helps is credited by name if they want to be.
This site is offered in 24 languages so that people who do not read English can take part. The words around the test were translated by a machine and have not yet been checked by a native speaker. Every translated page says so. If you read one of these languages, please use “Suggest a fix”: your wording is saved for a person to review and is never applied automatically. You can be credited for it, or stay anonymous.