Sources
| Source | Used for | Licence |
|---|---|---|
| Standard Korean Language Dictionary (표준국어대사전), National Institute of Korean Language, data of 5 June 2026 | origin type and Hanja of every word and counterpart; pronunciations; vocabulary statistics | CC BY-SA 2.0 KR |
| Urimalsaem (우리말샘), NIKL, data of 2 June 2026 | words not in the Standard Dictionary; dialect, archaic and North Korean labels; first-attested century; purification notes | CC BY-SA 2.0 KR |
| Basic Korean Dictionary (한국어기초사전), NIKL, data of 19 June 2026 | learner levels; pronunciations; quiz words and their English equivalents | CC BY-SA 2.0 KR |
| FrequencyWords, Korean list from OpenSubtitles 2018 (H. Dave) | rough subtitle counts (6.3 million words) | CC BY-SA 4.0 |
| Noto Sans KR, Noto Serif KR (Google) | Korean text, subset and served from this site | SIL Open Font License 1.1 |
| Supertonic 3 (Supertone Inc.) | synthesised Korean audio, made on our server at build time | OpenRAIL-M (model); MIT (software) |
| Whisper large-v3-turbo (OpenAI), via faster-whisper | recogniser check of the audio, on our server | MIT |
The full dictionary downloads of the Standard Dictionary and Urimalsaem require an account with the Institute, so we use the copy published by the korean-dict-nikl project, an unmodified mirror of the Institute's "download the whole dictionary" files. Dictionary example sentences quoted from publications are not under an open licence; we use none of them. All example sentences on this site are original.
How origins are checked
Every word and every counterpart is pinned to one dictionary entry: the Standard Dictionary entry with its homonym number, or, only when the word is not in the Standard Dictionary, a Urimalsaem entry. The site is rebuilt from that data, and the build fails if:
- a word marked native is not recorded as native (고유어) in its entry, or the entry gives it Chinese characters;
- a counterpart marked Sino-Korean is not 한자어, or its Hanja differ from the dictionary's;
- a word's origin is unrecorded: it has no pinned entry and is not explicitly marked as being in neither dictionary;
- a word marked "not in the dictionaries" turns out to be in the Standard Dictionary.
Of 280 words, 273 are native korean, 5 are not in the dictionaries, 2 are hybrid. Where the dictionaries call a word native but a foreign source has been seriously proposed, the page marks the origin as contested and gives both sides (4 words: 천둥, 장마, 사랑, 보라).
The vocabulary statistics in the guide count the Standard Dictionary's single-word headwords. Learner levels are joined to Standard Dictionary entries only when the match is unambiguous (same spelling and part of speech, and the same Hanja for Sino-Korean words); ambiguous homonyms are left out.
Pronunciation and romanization
Each word shows the dictionary's standard pronunciation (Standard Dictionary, else the Basic Korean Dictionary, else Urimalsaem). Our own engine applies the regular rules of standard pronunciation; checked against every native single-word entry in the Standard Dictionary with one recorded pronunciation, it gives the dictionary's form for 95.1% of 59,807 words, and most of the difference is tensing inside compounds (1,988 words), which depends on word structure, not spelling. On this site, 272 words follow the regular rules exactly, 1 has a lexical change the dictionary records, and 7 have no dictionary pronunciation (computed).
Romanization is the Revised Romanization of Korean, computed from the pronunciation without writing tensing, and tested on 42 examples from the official rules. The browser version of the engine (the sound-change tool) is checked against the build's on 937 words and sentences: 937 agree.
Audio
Piper, the open speech engine we use elsewhere, has only one Korean voice, trained on a dataset licensed for non-commercial use only, so we did not use it. The audio is made with Supertonic 3 (voice F1), on our own server when the site is built (a few clips come from ElevenLabs, below); every clip is served from this site, and nothing is fetched from anywhere else when you press play. We chose the voice by a recogniser round trip: each candidate voice read the same 30 words and sentences, a local Whisper model transcribed them, and we compared the transcript with the text.
| Voice | Exact (of 30) | Mean character error, words | Mean character error, sentences |
|---|---|---|---|
| Supertonic 3 F1 | 29 | 0.0 | 0.011 |
| Supertonic 3 F2 | 27 | 0.15 | 0.0 |
| Supertonic 3 M1 | 26 | 0.138 | 0.011 |
| Supertonic 3 M2 | 25 | 0.138 | 0.011 |
The same check runs on every published clip: 1013 clips, 885 transcribed exactly; mean character error 0.069 on single words and 0.018 on sentences. A single word is hard for any recogniser (it often writes the pronunciation, 같이 as 가치, or a commoner word that sounds alike), so the transcript is compared with both the spelling and the standard pronunciation. 27 clips scored badly. Where the paid ElevenLabs voice was understood and ours was not, we use its clip (6); 2 words have no audio because neither voice passed; the rest are single syllables the recogniser cannot judge out of context (it writes 입 as "1." and 피 as "P"), so they stay.
6 clips the open voice could not say clearly were made with ElevenLabs (eleven_multilingual_v2); 41 characters have been sent to that service in total, never by the weekly rebuild.
Subtitle counts
To give a rough sense of how often a rare word is really used, word pages for multi-syllable nouns and adverbs show how often the word occurs in the Korean OpenSubtitles 2018 frequency list (6.3 million words of film and TV subtitles). Korean attaches particles to words, so we count the word alone and with one common particle; verbs, adjectives and one-syllable words are not shown because their counts are polluted by unrelated words. The counts are a hint, not a measurement: subtitles under-represent written and literary Korean.
What the build checks
- origins and pins, as above;
- every example sentence contains its word;
- the romanization and pronunciation self-tests, and Python against JavaScript for the sound-change engine;
- every internal link and every related-word link resolves; no two words share a page address;
- the self-hosted font covers every Korean character on the site.