Write me four sentences of Korean to greet my Korean viewers with, and respell them in ordinary English letters so that I — who cannot read Hangul and have never heard the sounds — can say them out loud and be understood. Not IPA, not Revised Romanisation: what an English mouth has to see. Then I read it into Google Translate's microphone on camera and we all find out together.
This challenge scores itself, live, and that is the whole reason it exists. Paste the hangul into Google Translate to confirm it means what meaning claims. Then read read_aloud into Google Translate's Korean voice input and compare what it transcribes against the hangul it came from. That comparison is the score. Everything below is for explaining on camera why a column won or lost — it is not a substitute for doing the read-aloud.
Do the read-aloud for every model before you analyze any of them, and do it in one sitting, in the same voice, at the same pace. Your pronunciation improves noticeably across an afternoon of this, and if you read the columns in order the last model gets a better reader than the first one did.
Rule 1 is a grep, and it is brutal. Search every say and read_aloud for eo, eu, ae, ŏ, ŭ, ǔ and any IPA. Every hit is a failure of the one instruction the prompt gives first, longest and with examples. A model that writes an-nyeong-ha-se-yo has handed you Revised Romanisation — which is correct as romanisation and useless as instructions, because an English reader says "nee-ong" and "an-yong". The answer that does the job looks like ahn-nyuhng-hah-seh-yo. Expect at least one frontier model to fail this outright and to have failed it because it retrieved a standard instead of thinking about your mouth.
I present a YouTube channel in English. I want to open my next video by greeting my Korean viewers in Korean, out loud, in my own voice — and I do not speak a word of Korean. I cannot read Hangul. I have never heard most of these sounds.
What I can do is read English. That is the entire tool you have to work with: you write something in English letters, I say it into a microphone, and a Korean listener either understands me or does not.
Four sentences, in this order and in this order only, so I can learn them one at a time:
Keep each sentence short. I have to say this from memory on camera.
Respell it for an English mouth, not for a romanisation standard. Do not give me Revised Romanisation, McCune–Reischauer, Yale or IPA. eo, eu, ae, ŭ, ŏ, ʌ, ɯ are not instructions — I will read them as English and produce the wrong sound with total confidence. Write the syllables the way an English speaker would have to see them in order to say them: uh, oo, , , . If your respelling would survive being read aloud by someone who has never heard of Korea, it is right.
Rule 2 is where a Korean-built model should pull ahead. Korean is not pronounced as written, and respelling the spelling produces something that looks perfect and is wrong. Every trap below is a rule a Korean speaker applies without noticing. Check the ones that appear:
| Written | Actually said | What a failing column looks like |
|---|---|---|
합니다 -ㅂ니다 | ham-nee-da | hap-ni-da, -p-ni-da — the ㅂ nasalises to ㅁ before ㄴ. The highest-value check on the page: sentences 2 and 3 will almost certainly carry one of these endings. |
한국말 | han-goong-mal | han-gook-mal — ㄱ becomes ㅇ before ㅁ. |
한국에서 | han-goo-geh-suh | han-gook-eh-suh — the final consonant slides into the next syllable. |
못 합니다 | mo-tam-nee-da | mot-hap-ni-da — ㅅ/ㄷ + ㅎ fuse into an aspirated t. |
댓글 | det-ggeul | daet-geul — and the ㄱ tenses. Likely in sentence 4. |
좋아요 | jo-ah-yo | joh-ah-yo — ㅎ goes silent between vowels. |
같이 | ga-chi | ga-ti — ㅌ palatalises before ㅣ. |
환영 | hwa-nyuhng | hwan-yeong — the ㄴ links forward. |
A model that gets 합니다 right and everything else wrong still understood the assignment. A model that gets all of them right and wrote eo throughout understood Korean but not the person asking.
Then check the register, because this is the failure a native speaker hears first. You are a stranger addressing an audience of strangers. -요 endings (해요체) are warm and correct for a vlog; -습니다/-ㅂ니다 (합쇼체) is more formal and correct for a broadcast opening. Either is a good answer and the register field should say which it picked and why. Bare 반말 — sentences ending in a naked 해, 야, 어, or 안녕 standing alone as the whole greeting — is talking down to your entire audience, and it is a real failure however good the phonetics are. Mixing levels across the four sentences is sloppiness worth naming.
hard_sounds is the honesty axis. Two vowels genuinely do not exist in English and a model that does not warn you about them has not thought about the actual problem:
얼, 어. Nearest English is the u in bus. Models that respell it eo have set you up to fail.글, 는. There is no English equivalent at all: the u in pull with the lips deliberately unrounded, or a tight-throated uh. Any model with a usable description of this one is paying attention.Also worth credit: warning you that Korean has no English-style stress (English speakers hammer one syllable and it wrecks intelligibility more than any single vowel), that ㄹ between vowels is a flap like the tt in American butter, and that the tense consonants ㄲ ㄸ ㅃ ㅆ ㅉ are said with a tight throat and no puff of air. A model claiming the result will be "perfectly natural" is overselling and should be marked for it — the honest answer says a first-timer will sound foreign and aims for understood instead.
Rule 5 — did it trade a nicer sentence for a sayable one, and say so? A model that quietly wrote the elegant version with sounds you cannot produce optimized for reading well on the page rather than working in your mouth. That is the same error as rule 1, one level up.
Ranking: Google Translate recovers all four sentences from your reading, no romanisation artifacts, 합니다 nasalised, register named and correct > recovers most of it with one garbled sentence > correct Korean respelled as Revised Romanisation, i.e. a good translator that ignored the brief > wrong politeness level, or Hangul that does not mean what it claims.
eheeahRespell the sound, not the spelling. Korean is not pronounced the way it is written. Where the written form and the spoken form come apart, follow the spoken form every time — I am reading this to be heard, not to be transcribed back into Hangul. Getting this wrong is the single most likely way this whole thing fails, because the wrong answer looks completely reasonable on the page.
Break every syllable with a hyphen and keep the syllables even. Tell me if I should resist the English habit of hammering one syllable harder than the others.
Warn me about the sounds English does not have. Not a linguistics lecture — tell me what to do with my mouth. "Like the u in pull, but with your lips relaxed instead of rounded" is useful. "Close-back unrounded vowel" is not.
Pick the sayable version. If a nicer sentence needs a sound I would need three weeks to learn, write the plainer sentence that I can actually land. Say so if you made that trade.
Get the politeness level right. This is a stranger speaking to an audience of strangers on the internet. If you pitch it at the wrong level, a native speaker hears it instantly and it is the most embarrassing possible way to fail this.
You should know exactly what is about to happen to your answer, because it changes what a good answer is.
I am going to paste your Hangul into Google Translate to check it means what you told me it means. Then I am going to read your respelling aloud into Google Translate's Korean voice input and see what it thinks I said. What comes back out gets compared, on camera, to the Hangul you started from.
So this is not scored on how correct your romanisation is by any standard. It is scored on whether a machine that has never met me can recover your sentence from my English-speaker mouth reading your English letters.
Raw JSON, nothing else — no markdown fence, no preamble, no commentary afterwards. Exactly this shape:
{
"sentences": [
{
"hangul": "...",
"say": "...",
"meaning": "...",
"watch_out": ""
}
],
"read_aloud": "...",
"register": "...",
"hard_sounds": [
{ "sound": "...", "what_to_do": "..." }
]
}
say — the respelling, syllables hyphenated.meaning — a literal English gloss of that sentence, so I can check it against Google Translate.watch_out — only where this sentence has a trap in it. "" otherwise; do not pad.read_aloud — all four sentences as one continuous block of respelling, with / between sentences. This is the field I will actually read off the screen, so it has to stand alone.register — one sentence: what politeness level you chose, and why it is the right one for a stranger addressing an audience.hard_sounds — at most five. The sounds most likely to make me unintelligible, and what to do with my mouth.