Thai looks, at first, like a language a transcriber should sail through. The sounds are regular, the script is phonetic, and once you know the consonant and vowel signs you can read almost anything aloud. Then you try to write it down and the real problem shows up: Thai puts no spaces between words. A sentence is an unbroken run of characters, and the spaces you do see mark the end of a phrase or a sentence, not the gaps between individual words. That single fact turns Thai dictation into a word-segmentation problem: before the model can hand you correct text, it has to decide where each word starts and stops inside a continuous stream — and getting that boundary wrong produces text that is wrong even when every syllable was heard perfectly. Here's what separates Thai voice typing you can send from Thai voice typing you have to retype.
No spaces means the tool has to find the word boundaries
In a language that separates words with spaces, transcription and segmentation are the same act: you write the sounds, the gaps take care of themselves. Thai splits the two apart. The audio gives you a run of syllables; nothing in it tells you which syllables group into which word. ประโยคภาษาไทยเขียนติดกันแบบนี้ — a Thai sentence is written joined up like this — and the reader's eye does the segmenting silently. A dictation tool has to do that segmenting explicitly, and it has to do it right, because Thai has no visual cushion. In English a missing space produces something obviously broken you can't help but notice. In Thai the run of characters is supposed to be joined, so a wrong boundary doesn't look broken — it looks like a different, wrong sentence.
ตากลม: one string of letters, two different words
The reason segmentation is genuinely hard, not just fiddly, is that the same sequence of letters legitimately splits more than one way. The textbook example is ตากลม. Read it as ตา + กลม and it means "round eyes" (ตา eye, กลม round). Read the identical letters as ตาก + ลม and it means to air something out in the wind (ตาก to air or expose, ลม wind). Both are real Thai; only the surrounding meaning tells you which one was said. This is the crux of Thai dictation: the boundary can't be recovered from the sounds, because the sounds are identical. It has to be inferred from what the sentence is about. A tool that segments by acoustics alone is guessing, and half the time it will guess the reading you didn't mean — and hand it to you looking perfectly well-formed. Inferring the boundary from what the sentence means, not just its sounds, is how AI dictation works when it's built for a language written without spaces.
Tones carry meaning, and the script encodes them
Thai is tonal: the same base syllable is a different word at a different pitch, and the writing system spells that difference out through tone marks and consonant classes. The classic set is the syllable "maa" — มา with a mid tone means "to come," ม้า with a high tone means "horse," and หมา with a rising tone means "dog." These are not near-misses a reader forgives; they are unrelated words. Dictation has to map the pitch it heard onto the correct spelling, including the right tone mark and the right consonant, so that มา and ม้า and หมา don't get flattened into each other. When tone and segmentation both have to be resolved at once, you can see why "it heard the syllables" is nowhere near "it wrote the sentence." Thai shares this tonal burden with Vietnamese voice typing, where a single mark on a vowel decides the whole word.
Polite particles shift with politeness — and with the speaker
Thai ends a great many sentences with a politeness particle, and which one is correct depends on register and on the speaker. The standard polite sentence-final particles are ครับ, used by men, and ค่ะ, used by women — so the exact same request ends differently depending on who is speaking. Register moves them too: a casual message drops or softens the particle, a message to a client or a senior keeps it firmly in place, and the question form shifts again (a woman's polite statement ends ค่ะ but her polite question ends คะ). None of this is decoration. Using the wrong particle, or dropping it where it belongs, reads as either rude or oddly stiff. And the particle a raw transcript captures is whatever happened to come out of your mouth — which may not be the particle the finished message needs.
Why Thai dictation needs a rewrite pass, not just a transcript
Put those together and the shape of the problem is clear. A Thai transcript can nail every syllable and still be wrong three ways: the word boundaries fall in the wrong places, a tone gets spelled as its neighbor, and the ending particle fits how you spoke rather than who you're writing to. None of the three is fixable by transcribing more accurately, because none of them lives in the audio alone — they're decisions about meaning, about spelling convention, and about audience. That's the case for treating dictation as two steps — a twist on the practical ways to dictate on a Mac. First recognize the speech; then rewrite the result into Thai a reader would accept: boundaries set by what the sentence means, tones spelled correctly, and the polite particle chosen for the destination and the speaker. In English that second pass is a polish. In Thai it's the step that makes the output sendable.
Why "supports Thai" doesn't tell you much
A tool can list Thai on its feature page, transcribe one slow clean sentence in a quiet room, and still hand you a paragraph with the wrong word boundaries, a tone spelled as a different word, and a particle that suits a chat but not the email you were writing. The language count won't warn you about any of it. Dictate the way you actually work — a quick message to a colleague, then the same content as a note to a client, at your normal pace — and read what comes back. Is the text segmented into the words you meant? Are the tones the right words, not their neighbors? Does the ending particle match who you're sending it to? Would you send it, or fix it first? That is what one real session tells you, not a language count.