The Sound You Are Allowed to Keep

Gosh is God with everything after the first sound thrown away. So are golly, gad, gum and gummy. Heck is hell, darn is damn, frick is fuck, jeepers is Jesus. English has been manufacturing these for five centuries, and here are 76 of them measured phoneme by phoneme against the words they replace. The first consonant survives 52 times out of 76. Everything from the vowel onward survives not once.

This page is about words English will not say, so it prints them. All of them, unstarred, because a study of taboo vocabulary that flinches at the vocabulary is not a study.

One word, cut in two

An English syllable has a joint in it. Everything before the vowel is the onset; the vowel and everything after it is the rime, which is the part that has to match for two words to rhyme. cat and hat share a rime and differ in onset. That joint is where this whole page happens.

Take God. Take the euphemisms English has made for it and lay each one against the original, coloured at the joint. Most keep the left side. Not one keeps the right. The ones that change the left side are in the list too, and you can see which they are.

What that shows for one word, the whole set shows at once. Across all 76 pairs, the onset of the source word came through unchanged 52 times. The remainder came through unchanged no times at all.

The nearest thing to an exception is cornography for pornography, which is the only pair in the set that changes the onset and keeps almost everything else: 8 of its 10 phonemes are shared at the end, and what it swaps is the /p/ at the front. It does not count as keeping the rime, because the vowel right after the onset moves too. But it is the one word here built the other way round, and it exists.

Those two columns are the same 76 words asked two questions, so they can be compared directly. 52 pairs kept the onset and lost the remainder; 0 did the reverse. McNemar's exact test on that pair of counts gives p = 4.4 × 10^-16. Without the syllable analysis at all, on the blunter question of whether the first phoneme is the same phoneme: 59 of 76, against 37 of 76 for the last.

The cut that keeps this honest

Reading the table below showed a way that number could flatter itself. In a compound like motherfucker → mothertrucker the first consonant survives because the first word is never touched at all, and the same is true of shipload, forkhead and mollyfogging. So the same measures were run again on only the pairs whose source is a single syllable, where nothing can survive by accident because there is nothing standing in front of it. That subset was chosen after seeing the first result, which is worth knowing, so both are printed here and always will be.

The asymmetry is smaller on the honest cut and it does not go away: 23 of 37 against 0 of 37, p = 2.4 × 10^-7.

The other way this could have been an artifact

Not every word here is in the CMU Pronouncing Dictionary. Nobody has hand-transcribed jeepers, and CMUdict has never heard of cripes. So the missing ones come from Wiktionary's own printed IPA, or from espeak-ng's rules, and a comparison of phone strings across two transcription traditions can produce a difference that is about the traditions rather than about English. That is not a hypothetical: the first version of this page had fark as a perfect rime match for fuck, because Wiktionary prints a non-rhotic reading first, and it had cornography losing its /r/ to a Received Pronunciation reading. Both were transcription artifacts and both are gone: every pair on this page now takes both of its words from a single system, the best one that has both, and the 76 pairs divide 43 to CMUdict, 8 to Wiktionary IPA, 25 to espeak-ng.

The stronger version of the same test is to compute the whole thing three times, once inside each system on its own, over whatever subset that system can cover by itself. They agree.

espeak-ng can pronounce anything, so it is the only row that covers all 76 pairs, and it is a rule system rather than a record of how anyone speaks. CMUdict is hand-made and covers 43. The remainder column is 0, 0, 0.

The harder half: when English does not bend the word at all

There are two ways to mince an oath, and they are worth telling apart. You can bend the word until it is no longer the word: gosh, heck, frick, cripes, jeepers, none of which is English for anything else. Or you can swap in a word English already had, with a life of its own: shoot, sugar, fudge, flip, crud, snot, cow, Pete, Judas.

The bends are unsurprising. A deformation that kept nothing would not be recognisable as a substitute, so of course it keeps something. The swaps are the interesting case, because the choice was free. English was not editing shit into a new shape; it was reaching into its own dictionary for a word to stand in the same place. There are 125,930 words in the dictionary this page uses. It could have reached for any of them.

It reached for one that starts the same way 22 times out of 36. If the 36 substitutes had been drawn at random from that dictionary, the expected number sharing the taboo word's onset is 1.14. The exact probability of getting 22 or more by chance, summing over the real onset frequency of each source word, is 1.6 × 10^-25. Matching the first phoneme rather than the whole onset: 27 observed, 1.86 expected, p = 6.5 × 10^-29.

Restricting the dictionary to words of the same length as the substitute English actually chose, in case short taboo words can only take short replacements, moves the expectation from 1.14 to 1.25. And on the monosyllable-only cut, 9 of 20 against 0.58 expected, p = 8.4 × 10^-10.

The rime, meanwhile, is kept by 0 of the 36 swaps. Chance would have given 0.01. English is not avoiding the rime; the rime simply is not what it is holding on to.

Is this about taboo, or just about euphemism?

Here is the obvious objection, and it is a good one. Perhaps any word that stands in for another keeps its first sound, for reasons that have nothing to do with obscenity. English has thousands of euphemisms for things nobody is afraid to pronounce: pass away, restroom, big-boned, let go. If those keep the onset too, then the finding above is about substitution and the taboo is decoration.

Wiktionary has a category for those, and it is large. So the pairing rule was written as a program, with no judgement in it anywhere: take the first line of an entry that names a replaced word, take the first English word it links to that is not the entry itself or a piece of grammatical vocabulary, and keep the pair if both sides are one word with a pronunciation. That program was then run over both categories, 97 pairs from the minced oaths and 558 from the rest, 655 in all, with no hand audit on either side, so neither can be accused of having been curated into the answer.

Then the 655 pairs were split not by category but by the thing that actually varies: whether Wiktionary's own usage labels mark the replaced word as vulgar, obscene, offensive, profane or a slur. Not my judgement of what is taboo. Theirs, on that word's own page.

141 of 252 against 93 of 403. Fisher's exact test on that table gives p = 2.0 × 10^-17; on the stricter whole-onset measure, 131 of 252 against 97 of 403, p = 4.7 × 10^-13. The rime is kept by 2.4% and 1.2% of the two arms, which is to say by nobody.

So the effect is not euphemism. It is taboo. When what you are avoiding is a meaning, you are free to say it any way you like, and English mostly does: 23.1% is above the 5.1% two words would share by accident, largely because a euphemism is often built from the same stem, but it is not the pattern. When what you are avoiding is a word, its first sound comes with you.

Which makes sense, once you see it. A minced oath has a job: it must be understood, instantly, as standing in for the thing you did not say, while not being that thing. The onset is where a spoken word is recognised, before the rest of it has arrived. Keeping it is what makes the substitute legible. Losing everything after it is what makes it safe. The taboo lives in the rime.

All 76 pairs

Every row was read off the entry named in the last column, at the revision snapshotted in the repository. Green is the onset, red is the rime, and both words of a row always come from the same pronunciation source.

The published claim this replicates, and the check that keeps it from cheating

Lev-Ari and McKay argued in 2022 that swear words across languages under-use the approximants, the sounds /l/, /r/, /w/ and /j/, and offered English minced oaths as one line of evidence: the euphemisms, they found, contain more approximants than the words they replace. Their minced-oath set was twenty-four pairs. This one is 76, and it replicates: approximants rose in 25 pairs and fell in 9, a mean shift of 4.44 percentage points of the word, sign test p = 0.0090.

On its own that number would be worth very little, because minced oaths are slightly longer than their sources (4.45 phonemes to 4.99) and a longer word has more of everything. So the same paired test was fixed in advance to run on every class in the inventory, and all of them are printed here whatever they say. They do not all rise. Stops and fricatives fall, significantly, while nasals and vowels do not move. The word is not getting more of everything; it is trading obstruents for approximants.

What this rests on, and what it cannot say

The pair set is Wiktionary's category, which is what its editors have tagged and not a complete inventory of English minced oaths, so nothing here is a claim about all of them. Etymology is a discipline with real uncertainty and Wiktionary is a tertiary source; the hedged entries are dropped rather than guessed, which biases the set toward the well-established cases. The accent throughout is General American, which is what CMUdict encodes, and rhoticity matters here: darn keeps its /r/ in this reading and loses it in another. 10 pairs contain an r-coloured vowel, counted as a vowel and not as an approximant, which is a choice that could go the other way.

And the direction of the arrow is an inference, not a measurement. That gosh and God share an onset is a fact about two phone strings. That gosh was made from God is Wiktionary's etymology, quoted, and the whole result rests on those etymologies being right.