Skip to content
LanguageHelper

AI Diacritic Restorer

Paste text that was typed without accents — “cafe creme a Montreal”, “Viet Nam”, “Lodz” — and get the correctly accented version back.

Options

0

Output
Download

 

100% in your browser

About this tool

People often type without diacritics: on a foreign keyboard, on a phone without the right layout, or in systems that strip accents. The result is ambiguous and looks unprofessional. This tool reads each sentence in context with a compact multilingual language model and restores accents, cedillas, tildes, umlauts, breves, Vietnamese tone marks and Arabic harakat. It is designed to be safe: the model may only add marks. If its output changes a single letter or word, that passage is automatically restored with a dictionary of frequent words instead. Each language gets the restoration its writers actually need: ё in Russian, ї and ґ in Ukrainian, nukta and chandrabindu in Hindi, tone marks in Thai, Pinyin tones for Chinese, Hepburn macrons for Japanese romaji, McCune–Reischauer breves for Korean, and loanword accents (café, José) in English and Indonesian.

Like every tool on LanguageHelper, AI Diacritic Restorer runs entirely in your web browser. Your text is never uploaded, stored or used for training, which makes it safe for confidential documents, client data and personal messages. No account or installation is needed, and the tool keeps working offline once loaded.

Runs instantly with a built-in word-frequency dictionary — nothing to download. For context-aware results, choose an AI engine: its model (about 320 MB) downloads once and is then cached in your browser.

How it works

  1. Choose the text language or leave it on auto-detect.
  2. Paste your unaccented text into the input box.
  3. Click “Restore diacritics” for an instant dictionary result, or pick an AI engine for context-aware restoration (its model downloads once and is cached in your browser).
  4. Review the result, then copy it, download it, or open the Text Diff tool to see exactly what changed.

Use cases

  • Fix messages and e-mails typed on a phone or keyboard without the right layout.
  • Clean up names and addresses exported from systems that stripped accents.
  • Prepare correctly accented subtitles, captions and social posts.
  • Recover readable Vietnamese text written without tone marks.

Examples

French

Before

Ou passerez-vous l'ete prochain ? Il adore le cafe creme.

After

Où passerez-vous l'été prochain ? Il adore le café crème.
Vietnamese

Before

Bac si Nguyen song o Ha Noi.

After

Bác sĩ Nguyễn sống ở Hà Nội.

Tips for best results

For the best results, process complete sentences rather than isolated words: context is what lets the model choose between forms like “a” and “à” or “e” and “é”. Set the language explicitly when a text mixes languages. Very long documents are split into chunks automatically; if a chunk is returned unchanged, it was protected by the safety check and you can switch to the rules engine or edit it by hand.

Frequently asked questions

Which languages are supported?

All 20, each with its own job: accents in French, Spanish, Portuguese, Italian, Dutch and Romanian; umlauts in German; Polish and Turkish letters; Vietnamese tone marks; ё in Russian; ї and ґ in Ukrainian; tashkeel in Arabic; nukta and chandrabindu in Hindi; Thai tone marks; Pinyin tones; Japanese romaji macrons; Korean McCune–Reischauer breves; and loanword accents in English and Indonesian.

Is my text uploaded anywhere?

No. The model is downloaded to your browser and runs on your own CPU or GPU through WebAssembly or WebGPU. Your text stays in the page and is never sent to our servers or any third party.

Does it need to download anything?

Not by default. The instant engine uses a word-frequency dictionary of 10,000–25,000 words for your language (a few hundred KB at most). Only the optional AI engines download a model — about 320 MB for the fast one — which is then cached in IndexedDB, so later runs start in seconds and work offline.

Can the AI change my words?

Every chunk is checked after generation: with all marks removed, the output must be identical to your input. If the model altered, added or dropped anything, that chunk falls back to the dictionary engine, so your wording is preserved.

What does the instant (“Rules”) engine do?

It restores each word to its most common spelling, using frequency counts from millions of subtitle lines (tres → très, fuer → für, zolw → żółw, cok → çok), plus hand-checked lists (еще → ещё, ไม → ไม่) and exact converters for numbered Pinyin (ni3 hao3 → nǐ hǎo) and word-processor romaji (Toukyou → Tōkyō). Words whose spelling depends on context, such as French a/à or Spanish si/sí, are left as typed — that is what the AI engines are for, and they matter most for Vietnamese and Arabic.