The Prosody Workshop · a census

The Company Orange Keeps

Everyone knows nothing rhymes with orange. Almost nobody knows that nothing rhymes with have, and have is the ninth commonest word in the language. This is the exhaustive count: every English word that the dictionary leaves without a partner, the law that says which words those will be, and the short answer to why orange got famous and have did not.

The rule is fixed before anything is counted, and it is the ordinary one. Two words rhyme when they agree in sound from the last stressed vowel to the end, and when the consonants in front of that vowel differ. cat and hat rhyme. delight and light do not: they agree from the vowel onward, but they agree in front of it too, which makes them the same word twice rather than two words chiming. A word with no partner under that rule is what this page calls an orphan.

Those 48,614 words are the ones that two independent open sources both vouch for: SCOWL lists them as ordinary English (not a surname, not an abbreviation), and the Carnegie Mellon Pronouncing Dictionary knows how to say them. 18,747 of them have nothing to rhyme with. That is 38.6 per cent, and it is the first thing worth knowing about orange: being rhymeless is not rare. It is closer to normal than not.

The dial nobody hands you

"Nothing rhymes with orange" is a claim about a set, and no one who says it tells you which set. That is not pedantry, it is the whole difficulty. Run the question over the raw pronouncing dictionary and film rhymes with sovexportfilm, a Soviet film-export agency the dictionary happens to list, and chimney rhymes with zimny, a surname. Shrink the set far enough and everything is an orphan. The number moves, a lot, and the honest response is not to pick one set quietly.

So here is the dial. SCOWL grades English into size bands: band 10 is roughly the commonest few thousand words, and each band adds rarer ones. In SCOWL's own words, sizes "70 and below contain words found in most dictionaries", 80 adds "all the strange and unusual words people like to use in word games", and 95 is "just about every English word in existence and then some". This page's headline uses band 70, the largest set SCOWL itself will call dictionary English.

Every row: the census run again from scratch over that vocabulary. Band 70 is the row this page quotes.

Orphanhood is not a property of a word

Notice what the table does not do: it does not settle down. The share stays near two words in five whichever vocabulary you choose, and that is because each row changes two things at once. It changes which words are asked about and which words are allowed to answer. Separate them and something clearer appears.

Hold the question still. Take the 3,961 commonest words of English, the ones in SCOWL's smallest band, and ask only about those. Now grow the pool of words permitted to rhyme with them.

The count falls by more than half. If the only words you know are the commonest few thousand, 1,599 of them have nothing to rhyme with. Learn the rest of the dictionary and 887 of those find a partner, leaving 712. So the majority of the rhymelessness that a speaker of ordinary English actually experiences is not a fact about English at all. It is a fact about the size of their vocabulary. An orphan is not a word on its own; it is a word and a vocabulary, and the sentence "nothing rhymes with X" always has an unstated second half: that I know.

The law

What decides it, then? Almost entirely one thing, and it is not obscurity or foreignness or spelling. It is how much word there is after the stress. A rhyme has to match every sound from the stressed vowel onward, so the longer that stretch, the more exact agreement a partner has to supply, and the fewer candidates survive.

Measure it. Here is the share of words that are orphans, cut by the length of the stretch they demand, in phonemes.

It is close to a staircase. A word whose rhyme is two sounds long is essentially always partnered. A word whose rhyme is eleven sounds long never is. And somewhere in the middle the coin flips.

The orphans that deserve their fame

If orange is unremarkable, which words are not? The ones that are orphaned where orphanhood is nearly unheard of: short rhymes, where the base rate is well under one per cent. There are 29 words in this census whose rhyme is only two sounds long and which still have no partner. They are the genuine article, and with one exception nobody has ever put them on a list.

Read that line again. have. Not a Scots dialect word or a botanical term: the ninth commonest word in written English by the Oxford English Corpus's count of two billion words, and the eighth by the Corpus of Contemporary American English. Its rhyme is two phonemes long, and there is nothing in the dictionary to match it. (That ranking is those corpora's, not this census's: SCOWL grades commonness in bands rather than ranks, and its smallest band holds 3,961 words, so nothing here could have told you have was ninth. It is cited, not computed.) Nor with oblige, beige, rouge, prestige, mauve, fugue, gouge, babe, duo or trio.

And there is a joke buried in the first two entries. have is HH AE V. So is halve. They are homophones, which means they share the consonant in front of the vowel, which means they are the same word twice and not a rhyme at all. The one word in English that sounds exactly like have is the one word that is disqualified from rhyming with it.

The famous list, adjudicated

Here is the folk canon, put through the same rule as everything else. The rhyme column is the stretch that a partner would have to match; base is the share of words with a rhyme that length that turn out to be orphans, which is the fair yardstick for how surprising each verdict is.

Three answers that are not answers

Sporange does not rhyme with orange. This is the standard rebuttal and it does not survive contact with a dictionary. Sporange is real, a dated variant of sporangium, a spore case. But Merriam-Webster gives it as spə-ˈranj, the OED as /spɒˈrændʒ/, and Wiktionary's British and American entries both stress the second syllable. Stressed there, its rhyme is -anj, which puts it with range and hinge, not with orange. Wiktionary does list a first-syllable, reduced-vowel pronunciation that would rhyme, and labels it humorous. That is a pronunciation which appears to exist because people want the rhyme. Merriam-Webster's own verdict on it is that the word is "all but useless as a rhyming word with orange in the realm of poetry".

Algorithm does not rhyme with rhythm, and billionth does not rhyme with month. Both fail the same way, and it is the way most offered rhymes fail: stress. algorithm is AE1 L G ER0 IH2 DH AH0 M, with its primary stress on the first syllable, so its rhyme is the whole word from AE onward, not the -ithm tail your eye picks out. Same for billionth: stressed BIL, so it belongs with million's family and never with month. A rhyme is anchored at the stress, and the eye does not see stress.

The bug in the standard method

The rule above has a trap in it, and this census fell into it before it climbed out. To find the consonants in front of the stressed vowel you walk backwards from that vowel and collect what you meet. For an ordinary word that is right. For a compound it is not, and the pronouncing dictionary makes compounds visible by giving them two primary stresses:

The walk crossed a syllable boundary and swallowed the K that ends quick. So the onsets look different, and the method reports silver and quicksilver as a perfect rhyme. No ear accepts that: you can hear the whole word silver sitting inside it.

The fix is not a special case for compounds. It is the syllabification rule the backwards walk was missing. English does not permit /ks/ at the start of a syllable, and under the Maximal Onset Principle the true onset is the longest run of consonants that English actually allows there. For quicksilver that is S, identical to silver's, which makes the pair the same word twice, which is what it is. The page states its whole legal-onset inventory, cluster by cluster, each with a word that carries it, and the verifier looks every one of them up in the dictionary rather than taking the list on trust.

Applying the correction costs 719 words a rhyme they never had. It also turned up a live error in this project's own published rhyme lab, which used the uncorrected walk and therefore called silver and quicksilver perfect. That page has been corrected; the note is in the apparatus at the foot of this one.

Crossing the Atlantic

English without the r ought to rescue a crowd of these words, because dropping coda /r/ merges rhyme classes that General American keeps apart. And it can only ever help: merging classes never splits one, so the orphan count can go down and never up. That is a fact about the map, and the verifier proves it rather than asserting it.

It rescues 292 words out of 18,747. That is almost nothing, and the reason is exact, and funny.

Every famous transatlantic merger produces a homophone. And a homophone shares your onset, so it is the same word twice, so it is exactly the thing that cannot rescue you. Crossing the Atlantic hands you twins, not partners.

The hole in the floor

Now the part that matters most, because it is where this census is weakest and the weakness is measurable.

Every number above rests on the Carnegie Mellon Pronouncing Dictionary, a hand-built lexicon of 126,052 headwords. SCOWL knows 425,058 words. So there are hundreds of thousands of English words with no CMU pronunciation, and every one of them is a rhyme this census cannot see. That is not a footnote. It decides the case this page opened with:

Calve rhymes with have. It is an ordinary word. The census called have an orphan only because the lexicon it trusts has a hole in exactly that spot. The same is true of loaf, whose rhyme oaf is also missing, and of the classic answers to the famous words: chilver (a ewe lamb, Merriam-Webster, "now dialectal, England") really does rhyme with silver, and curple and hirple (both Merriam-Webster, both Scottish) really do rhyme with purple. Every one of them is absent from the pronouncing dictionary. The famous escape hatches all live outside the book that would have to adjudicate them.

So how big is the hole?

To measure it you need pronunciations for words the dictionary does not have, which means a second instrument. eSpeak NG is a rule-based letter-to-sound system, so it will pronounce anything, which makes it right for measuring the gap and wrong for setting headline numbers. It is used only as a probe, and it is calibrated before it is trusted: on the 48,610 words both sources can pronounce, the two agree on the orphan verdict 92 per cent of the time, and produce a character-for-character identical phone string 54.4 per cent of the time. Both numbers are printed whether or not they flatter the probe.

So somewhere between one orphan in twenty and two in five is an artifact of dictionary coverage rather than a fact about English, depending on how obscure a rescuer you will accept. And the rescuers are mostly not obscure at all: 31.8 per cent of them, at the band-70 setting, are a regular inflection of a word the dictionary already contains. It has anoint but not anointing, roast but not roastings. The hole is made of ordinary morphology, not of Scots dialect.

Checking the probe by hand

A count nobody has checked is a guess with a decimal point on it. So a deterministic sample of 27 of those claimed rescues was pulled out (every Nth of the sorted list, so the draw spans the alphabet) and every pair was looked up by hand against Merriam-Webster, Collins, Wiktionary and Dictionary.com: is the rescuing word real, how is it actually said, and does it truly rhyme?

The failures are the instructive part, and they are all the same kind of failure. Not one of them is the census misapplying the rule: every one is eSpeak mispronouncing a word. It gives mitral the wrong vowel. It puts the stress on the wrong syllable of adnate and stagnate. It reads parget with a hard g when the word is /ˈpɑrdʒɪt/. And into intussusception it inserts a /k/ that is not in the word, which is what made that pair look like a rescue at all: with the real pronunciation the two onsets match, so it is identity, which the rule already excludes.

Stress and invented segments. That is exactly the failure mode a rule-based letter-to-sound system has, and it is exactly the failure mode behind the sporange claim higher up this page: the eye reads a tail off the spelling and puts the stress in the wrong place. The probe and the folk tradition make the same mistake, which is worth knowing about both of them.

The honest headline is therefore a range, not a number. Of the 48,614 words this census can see, 18,747 have no rhyme in it, which is 38.6 per cent. Discount the rescue counts by the audited hit rate and the figure falls to about 32.8 per cent if you will accept any dictionary word as a rescuer, and toward 27 per cent if you will accept anything SCOWL lists at all. This page cannot pin it tighter than that. Doing so would need an open pronouncing dictionary that covers the whole of English, and there is not one.

Look up any word

The instrument at the top of the page is the census itself, not a summary of it: the 48,614-word index, the 24,777 rhyme classes, and the same rule. Give it a word and it will tell you whether it is an orphan, what its rhyme is, who shares it, how surprising the verdict is given its length, and, when there is no rhyme at all, the nearest thing in the language and how far away that is.

Back to the instrument ↑