Korean audio on Learn Korean Hangul is generated with a text-to-speech engine, not recorded by a person, and Min-ji Park, a Korean language teacher who reviews the site's Korean, listens to and approves each clip before the page that uses it is published. This page explains how clips are made, checked and corrected.
Where the Korean Audio Comes From
The audio comes from Chirp 3 HD, a set of text-to-speech voices in Google Cloud Text-to-Speech. It is a cloud service, not software the site runs itself: the site sends Korean text to Google's API and receives the finished speech. Since 10 October 2026 it has been the default engine, replacing an open-weight engine the site had planned to run on its own machines. Single clips use the Korean voice named Kore, and two-person dialogues add a second Korean voice, Charon, for the other speaker.
Two facts about this engine matter to a learner. First, because it is a cloud service, the Korean text of each clip is sent to Google to be spoken. That text is the lesson's own Korean, not anything about you: no learner data is sent, and nothing is sent when you play a clip. Second, a voice that sounds fluent is not proof that it teaches Korean correctly. Whether its tense and aspirated sounds, final consonants and sound changes are right for a learner is unverified until a Korean speaker listens.
Next, the pipeline shows how one clip is produced.
How Is a Clip Produced?
A clip is produced in six steps, all before you visit the page. The engine never runs while you play audio, so the page only plays a finished file.
- List the item. Each Korean string that needs a clip is listed on its page with a stable identifier and a context, such as "letter name" or "example sentence".
- Prepare the input. The text sent to the engine is written in Hangul, with numbers spelled out in words. Letter names are sent as names, such as 기역, never as the bare symbol ㄱ.
- Generate. A script on the site's server sends the text to Google's API with a recorded voice and speaking rate. It keeps under a request rate limit and slows or pauses when the server is busy. The master file is stored as uncompressed audio (24 kHz mono WAV) in private storage.
- Check. Automated checks confirm the file, length and loudness. Min-ji Park, the site's Korean reviewer, then listens to the clip and approves it or sends it back. She reviewed and approved every clip on 15 October 2026.
- Compress. Public copies are made in a smaller web format, with long cache times so a replayed clip loads quickly. The exact web format has not been chosen yet.
- Record. The clip's record stores its review status, the reviewer and the date, and a page can be published only with approved clips.
Each clip's record also stores the voice and the text the engine was given, so any clip can be traced. If a clip's text, voice or speaking rate changes, the clip is generated again, and the new version starts unreviewed.
Why Can the Audio Differ from the Spelling?
The site sends the engine plain text only, with no phonetic markup, so the control it uses is the text itself. The site therefore separates the display text you read from the pronunciation input the engine hears.
Take 같이. The spelling is 같이, but the standard pronunciation is [가치] (gachi), because a ㅌ before 이 softens to ㅊ and links to the next syllable. The page shows 같이, and the engine is given 가치. A clip built this way matches the sound without changing the spelling you learn.
This respelling is a fix of last resort, applied only to items that failed review. If it does not work, the fallback is a different voice, a different engine or a human recording. Next, the standard the clips are measured against is named.
Pronunciation Standard behind Each Clip
Clips target standard Seoul pronunciation as set out in the Standard Pronunciation rules (표준 발음법) of the National Institute of Korean Language (국립국어원). The language standards page lists these rules beside the spelling and romanization standards. The rules themselves are taught in the Korean pronunciation hub.
A synthetic voice is not a person from Seoul, so the claim is about the target, not the result. The clip is meant to follow the rule. Whether it does is what review decides.
What Do the Audio Review Statuses Mean?
Each clip's record holds one review status, and the status states what has actually been done. Only approved clips appear on a published page.
| Status | What it means | Can the clip be published? |
|---|---|---|
| Unreviewed | Generated by the engine and passed the automated checks; no one has listened yet | No |
| Needs fix | A reviewer heard a problem that a respelled input, another voice or new settings may solve | No |
| Rejected | The clip cannot be fixed and is replaced, for example by a human recording | No |
| Approved | Min-ji Park listened to the clip and approved it | Yes |
The status cannot be set by hand to a stronger state than the review that happened. An approved clip is still machine-generated speech, and the site says so. Her role and background are described on the authors and reviewers page. Next, the known limits are listed honestly.
Known Limits of Synthetic Korean Speech
Synthetic Korean fails in predictable ways. The site tests for these first, because learners depend most on exactness here:
- Tensification after ㄹ in Sino-Korean words. 발달 (development) is pronounced [발딸]: romanized baldal, but said with a tense ㄸ.
- N-insertion. 한여름 is pronounced [한녀름] (hannyeoreum), with a ㄴ sound that is not in the spelling.
- Long and short vowels. Dictionaries mark length in pairs such as 눈 (eye) and 눈: (snow), but spelling does not, and modern speech often drops the contrast.
- Isolated letters and syllables. An engine trained on sentences can truncate or distort a single sound.
- Letter names. The name 시옷 is pronounced [시옫], because final ㅅ is pronounced as ㄷ.
- The vowel ㅢ. It has several accepted pronunciations depending on the word, such as 의 as [의] or [에] when it is a particle.
The three clips above are the tensified, n-inserted and letter-name cases, in that order. Treat them as test cases rather than models.
Pilot Test before Audio Goes to Scale
Before audio is produced at scale, a pilot of at least 100 items is planned. It covers 14 letter names, 12 vowels, 8 syllables, 14 plain, tense and aspirated contrasts, 8 final-consonant cases, 10 sound changes, words, sentences, dialogues, numbers and speech levels. Min-ji Park rates each clip blind for correct pronunciation, audible contrast and naturalness, against a pass mark agreed with the site owner before the pilot starts. Items that fail are fixed or replaced, and large-scale production follows the pilot, not the other way round.
Next, how to use the audio well.
How to Use the Audio Well
Audio helps most when you treat it as a listening and shadowing tool. Tap to play, since nothing plays automatically. Use the slow setting for a clip you cannot follow, which plays the same file at 0.8 times speed, or a separately generated slow clip if the slowed version sounds distorted. Then repeat the clip aloud and compare by ear.
If a clip sounds wrong to you, trust that signal and check a dictionary. The site does not listen to you and gives no pronunciation score.
Reporting a Wrong Clip
Send the page address, the Korean word and what you heard to the corrections page process. A reported clip is checked against the standard, then fixed by respelling the input, changing the voice or setting, or replacing it. The fixed page carries a dated note. Text-to-speech is one of the AI uses listed on how we use AI, and the About Learn Korean Hangul page gives the wider picture.
Frequently Asked Questions
Why not use human recordings? Human recordings are the better teaching standard, but recording thousands of items takes far more speaker time than reviewing them. Synthetic audio, approved clip by clip by Min-ji Park, the site's Korean reviewer, makes wide coverage possible. The plan lets a human recording replace a generated clip for the same item, starting with letter names and vowels.
Is the audio good enough to copy for my accent? Use it as a guide, not as a model of accent. The clips target standard Seoul pronunciation, and each must be approved by a reviewer before it is published, but synthetic voices can still flatten rhythm and pitch. For fine detail, such as the plain, tense and aspirated contrast, compare with a dictionary's audio or a teacher.
Why do letter names sometimes sound odd? Letter names end in consonants that change when spoken alone. The name 시옷 is pronounced [시옫], because a final ㅅ becomes a ㄷ sound. Engines trained on sentences can also handle single words poorly. Letter names are marked for priority human review, and an odd clip should be reported.
Does the site score my pronunciation? No. The site does not record your voice, listen to you or give a pronunciation score, and it will not claim to unless a scoring method has been validated. The audio is for listening and shadowing only: play a clip, repeat it aloud, and compare by ear.
Can I download the audio? No download option is promised. Clips are played from the page, and the site is built so that your browser caches a clip after the first play. If downloads are ever offered, this page will say which clips and under which terms.