How Often the Thai Tone Rules Actually Work
TL;DR: The Thai tone rules have a reputation for being riddled with exceptions, and on this corpus they are not. Applied to 1,844 single-syllable words, consonant class plus syllable type plus tone mark predicts the right tone 1,812 times - 98.3 percent, and the same figure when each word is weighted by how often it occurs in the Thai National Corpus. There are 32 exceptions. Twenty-eight of them are English loanwords, and one of the remaining four is ก็.
The rules have a worse reputation than they deserve
Every learner meets the tone rules in the same order. First the promise: the spelling tells you the tone, so you never have to memorise one. Then the table, which has nine cells and needs three inputs. Then, almost immediately, somebody in a forum says the rules are full of exceptions and you will end up learning the tones word by word anyway.
That last claim is the one worth checking, because it changes what you should do with your first month. If the rules hold, learning nine cells replaces memorising thousands of tones. If they leak badly, the cells are a detour. So the question is not whether exceptions exist. It is how many, and whether you will meet them.
98.3 percent, counted
Phayan ships a word list of 2,249 Thai words, each carrying its own IPA, its consonant class, whether its syllable is live or dead, and the tone it is actually pronounced with. That is everything the tone rules take as input and the answer they are supposed to produce, which makes the list a test set for the rules themselves.
The test was restricted to the 1,844 words that are a single syllable with a single consonant class, because those are the words where the rule has exactly one derivation to make and no ambiguity about which consonant is the initial. For each one, the standard derivation was applied - class, then live or dead, then the tone mark if there is one - and the result compared against the stored tone.
It got 1,812 of them right. That is 98.3 percent, and weighting each word by its Thai National Corpus count gives the same 98.3 percent across 18,646,600 word tokens. Whichever way you count, the rules are not leaking.
Phayan is an app for learning to read Thai script, and it is built on that result rather than around it: the levels go letter by letter from zero, and the tone rule arrives once enough of one consonant class is in place for the derivation to be usable instead of theoretical. Every word in the app carries audio, and your progress is kept without an account.
The exceptions are mostly English
Thirty-two words broke the rule. Twenty-eight of them are loanwords or the Thai names of Latin letters, and twenty-five of the thirty-two are pronounced with a high tone where the derivation predicts low or falling.
| Word | Means | Rule says | Actually |
|---|---|---|---|
| เมตร | metre | falling | high |
| บัส | bus | low | high |
| ซอส | sauce | falling | high |
| เคส | case | mid | high |
| เชฟ | chef | falling | high |
| เอฟ | the letter F | mid | high |
There is a reason they cluster like that. A borrowed word is spelled to match the sound of the source word, and the speaker then says it the way they heard it rather than the way the page derives it. High tone is where those words land, often enough that it works as a rule of its own: if a word looks like an English word you already know, expect a high tone and stop deriving. That covers เฟก, เซฟ, บอส, ชีส, พีช, โคช, เทป, แอป and most of the rest of the list.
This is not a small category and it is growing, which is the honest caveat on the 98.3 percent. A word list drawn from news and social media in 2030 will carry more loans than this one does, and every one of them is a word whose tone the spelling does not settle.
And then there is ก็
Four exceptions are native Thai, and one of them matters more than the other thirty-one put together. ก็ is a mid-class consonant with a short vowel and no tone mark, so the derivation says low tone. It is pronounced with a falling tone. It also occurs 299,255 times in the Thai National Corpus - the whole set of exceptions is 1.7 percent of the weighted tokens here, and ก็ by itself is 1.6 of that 1.7.
So the practical version of this whole article is short. Learn the rules, learn that borrowed words tend to go high, and learn ก็ as a one-word exception the way you learned it in English that "one" does not rhyme with "bone". That is the exception list. The other three native oddities - เมือ, แซบ and ช่ะ - are a dialect form, a known misspelling of แซ่บ and a particle, and between them they account for a rounding error.
A tone mark is more reliable than no tone mark
The count splits cleanly in a way that surprised me. Of the 1,844 words, 297 carry one of the four tone marks and 1,547 carry none. The marked ones derive correctly 99.7 percent of the time; the unmarked ones, 98.0 percent. A tone mark is the more trustworthy signal, not the more complicated one.
That runs against how the marks are usually taught, as a late complication on top of an already awkward table. They are the opposite: a mark names the tone almost unconditionally once you know the class, and mai tri and mai chattawa only ever appear on mid-class consonants, so they name it outright. The uncertainty lives in the unmarked words, which is also where most of the language is - unmarked words are 68 percent of the running text measured here, and only 329 of the 2,249 words in the list carry a mark at all.
| Spelling | Words | Share of running text | Rule correct |
|---|---|---|---|
| No tone mark | 1,547 | 68.0% | 98.0% |
| With a tone mark | 297 | 32.0% | 99.7% |
If you are choosing what to drill, that is the answer: the unmarked live and dead distinction is worth more practice than the marks are. The mechanism behind both is in the tone rules themselves, and which letters belong to which class is in the three consonant classes.
What the count found in our own word list
The first pass came out at 97.1 percent, and the 53 failures had a shape: 34 of them contained the vowel -ัว. The word list types that syllable as short, which makes วัว a dead syllable and predicts a high tone; -ัว is long, วัว is live, and the tone is mid, which is what the list stores. Retyping those 34 and running the count again is what gives the 98.3 percent above.
Worth being precise about what that is and is not. The tone stored against each word is derived from the word's own IPA, so the tones the app teaches are not affected - it is the live-or-dead label on 34 of the -ัว words that is wrong. It is recorded here because the 98.3 percent is not reproducible without it, and a number you cannot reproduce is not a measurement.
What to do with this
- Learn the nine cells. On the evidence here they settle the tone of roughly forty-nine words in fifty, which is not a shortcut you can get any other way.
- Treat a borrowed word as its own category. If you recognise the English inside it, expect high tone and move on rather than deriving and doubting yourself.
- Memorise ก็ and stop worrying about the rest of the exception list, because it is statistically almost empty.
- Spend the practice on live versus dead, not on the four marks. That is where the remaining 2 percent of error actually sits, and it is 68 percent of what you will read.
The counting method here is the same one used for which consonants actually appear, and the two results point the same way: the script is more regular than its reputation, and the irregularity that remains is concentrated in a handful of items you can name.
Frequently asked questions
What is the best way to learn the Thai tone rules?
Learn the consonant classes first, then the live and dead distinction, and only then the four marks - in that order, because the first two decide the tone of 68 percent of the words you will read and the marks are nearly deterministic once the class is known. Phayan sequences its levels that way and introduces the marks at level four, after enough of the mid class is in place for the derivation to be usable. Anki with a hand-made deck is the closest do-it-yourself version, and it will drill the nine cells without ever making you derive one under time pressure.
How do I know when a Thai word breaks the tone rules?
Nearly always because it is borrowed. Twenty-eight of the 32 exceptions found in this 2,249-word list are loanwords or the Thai names of Latin letters, and twenty-five of the 32 are pronounced high where the rule predicts low or falling. If the word does not look borrowed and does not derive, the realistic possibility is that you have the consonant class or the syllable type wrong. Phayan drills those two inputs before the marks for exactly that reason.
Is it safe to rely on the tone rules instead of memorising tones?
Yes, on this measurement: the derivation is right 98.3 percent of the time, both per word and weighted by how often the words occur. What it gives you is the tone the spelling implies, which is a different skill from hearing or producing that tone. Phayan plays audio for every word so the two can be practised together, and the audio is synthesised rather than recorded, so treat a native speaker or a teacher as the authority on how a tone should sound. ThaiPod101 takes the opposite route and teaches the tones by ear without asking you to decode at all.
What is the difference between a tone rule and a tone mark?
The rule is the whole derivation - consonant class, syllable type and mark together; the mark is only one of its three inputs, and most Thai words do not have one. Only 329 of the 2,249 words in this list carry a mark, so a reader who has learned the four marks and skipped the classes has learned the smaller half of the system. Phayan introduces the marks at level four, once the classes they depend on are already in place.
So do I really have to memorise ก็ separately?
Yes, and it is the one word on this list worth the trouble. It derives to a low tone and is said with a falling one, and it turns up 299,255 times in the Thai National Corpus, which makes it more common than almost anything else you will misread. Phayan teaches it as vocabulary rather than as a derivation, which is the honest way round: learn it as a word, not as a rule, and the rest of the exception list can wait until you trip over it.