← All models
    Diacritic restoration

    Put the accents and tone marks back, in 11 languages

    People type without diacritics because phone keyboards make them slow, and old databases strip them. DiacNet restores Yorùbá tone marks, Vietnamese accents, Polish and Turkish letters and Arabic vowels, without rewriting your text.

    $1 of free credit · open weights on Hugging Face

    Text typed without marks in Yorùbá, Vietnamese, Polish and Turkish, restored with its diacritics highlighted
    11
    Languages, including Arabic in two styles
    0.0117
    Mean error rate (diacnet-2.0)
    2.5×
    Lower Yorùbá error than the previous version
    $1.00
    Per million characters on the API
    / What it does

    11 languages

    Yorùbá, Igbo, Hausa, Vietnamese, Polish, Turkish, Portuguese, Spanish, French, Italian and Arabic.

    Never rewrites your text

    Alignment keeps every letter you typed and takes only the marks from the model.

    Meaning hints

    Tell it which meaning you want for an ambiguous word: Vietnamese ambiguous-word accuracy rises from 0.63 to 0.90.

    Arabic in two styles

    Full vowelling with case endings, or the lighter style most modern Arabic uses.

    Runs on a CPU

    diactag-2.0 restores a typical sentence in about 8 ms on a laptop, and by design can never change a letter.

    Open weights

    Run it on the API, or download it and process archives in your own environment.

    / Live output

    See it restore marks

    Pick a language. The output is exactly what diacnet-2.0 returned on our API; the restored marks are highlighted.

    Typed without marks

    O so fun ara re pe oun ko ni isoro kankan.

    diacnet-2.0

    Ó sọ fún ara rẹ̀ pé òun kò ní ìṣòro kankan.

    Text typed without marks, as people do on phone keyboards.

    / Benchmarks

    A big step for Yorùbá

    Published results for diacnet-2.0.

    Yorùbá error rate

    Character-level error, lower is better. Shorter bar is better.

    • diacnet-2.00.070
    • diacnet-1.10.174

    Vietnamese ambiguous words

    Accuracy on words whose marks depend on meaning.

    • diacnet-2.0 with meaning hints0.900
    • diacnet-2.0 without hints0.630
    Model card and evaluation details
    / Built for

    Text-to-speech and voice

    Tone marks decide pronunciation in Yorùbá and Vietnamese: restore them before speech synthesis.

    Search and archives

    Restore marks in old databases and OCR output so search and display read properly.

    Writing tools and keyboards

    Let people type quickly without marks and give them properly written text back.

    / By sector

    Where it fits

    DiacNet puts back the tone marks, accents and vowel marks that people leave out when they type: Yorùbá, Igbo, Vietnamese, Arabic and seven more. Missing marks change meaning, break search and make text-to-speech mispronounce, so restoring them matters in any sector that publishes, speaks or searches text in these languages.

    Publishing and media

    Readers notice missing tone marks; editors don’t have time to add them by hand.

    • →Restore Yorùbá tone marks and European accents in articles, books and subtitles before publication, with diacnet-2.0, the most accurate of our models on 8 of 10 languages.
    • →Give a meaning hint for ambiguous words, such as [g: ogun=war], and the right marks follow.

    WatchRestoration is very good, not perfect: in Yorùbá about 7 characters in 100 can still be off, and only 12% of sentences come back exactly right. Have a fluent editor proofread anything published.

    Speech and voice products

    • →Restore marks before text-to-speech, so ògùn (medicine) isn’t read as ogun (war).
    • →Full Arabic vowelling with case endings for speech synthesis and teaching, or the lighter no-case-endings style most modern Arabic uses.

    Education

    • →Prepare fully marked reading material for learners of Yorùbá, Igbo or Arabic.
    • →Turn unmarked texts students type into marked versions they can compare against.

    WatchFor teaching material, a teacher should check the output: learners copy what they see.

    Search and data

    • →Normalise queries and documents so a search typed without marks still finds marked text, and the other way round.
    • →Restore marks across corpora before training speech, translation or language models on them.

    Archives, legal and records

    Where letters must never change, use diactag-2.0.

    • →diactag-2.0 copies every base character and only adds marks, so it cannot alter your text: letter safety 1.0000 by design.
    • →Its per-character confidence flags unsure marks for an archivist to review, and the 38.6 MB ONNX export runs on ordinary CPUs.

    A production pattern that works across sectors

    1. 01Choose the model: diacnet-2.0 for the best accuracy, diacnet-mini-2.0 for lower cost, diactag-2.0 when letters must never change or you need CPU.
    2. 02Keep alignment on, so every letter you typed is kept and only the marks come from the model.
    3. 03Add meaning hints for words you know are ambiguous in your domain.
    4. 04Use per-character confidence (diactag) or a proofreader’s sample to decide what needs review.
    5. 05Keep corrections and check accuracy per language over time, especially Yorùbá, where the task is hardest.
    / The lineup

    The DiacNet models

    DIACRITICSMOST ACCURATE

    diacnet-2.0

    Put the accents and tone marks back, in 11 languages.

    11
    Languages
    0.0117
    Mean error rate
    2.5×
    Lower Yorùbá error than 1.1
    582M
    Parameters
    Best for
    • Publishing-quality restoration in Yorùbá and the European languages, where accuracy matters most.
    • Text with ambiguous words, where a meaning hint settles the right marks.
    • Messy input where you also want typos corrected along the way.
    DIACRITICS

    diacnet-mini-2.0

    diacnet-2.0’s features at half the size.

    11
    Languages
    0.0141
    Mean error rate
    2.1×
    Lower Yorùbá error than 1.1
    300M
    Parameters
    Best for
    • Large backlogs and busy products where serving cost matters.
    • Polish, Portuguese, Spanish, French and Italian, where it outperforms diactag-2.0.
    • Teams who want diacnet-2.0’s features on smaller GPUs.
    DIACRITICS

    diactag-2.0

    Diacritics that can never change your letters.

    11
    Languages
    1.0000
    Letter safety, by design
    ~8 ms
    Per sentence on a laptop CPU
    37.9M
    Parameters
    Best for
    • Records, legal text and archives, where letters must never change.
    • CPU-only and on-device serving through the int8 ONNX export.
    • Modern Standard Arabic, Igbo and Vietnamese, its strongest languages.
    / Side by side

    Compare the DiacNet models

    diacnet-2.0 →diacnet-mini-2.0 →diactag-2.0 →
    Parameters582M300M37.9M
    How it worksRewrites the textRewrites the textTags each character
    Mean error rate, 10 languages0.01170.01410.0130
    Yorùbá error rate0.06980.08150.0787
    Strongest onYorùbá and European languagesClose to 2.0 at half the sizeIgbo, Vietnamese, Modern Standard Arabic
    ExtrasMeaning hints, typo correctionMeaning hints, typo correctionCPU-ready ONNX, never alters letters
    / Choosing

    Which one should you use?

    • Best diacritic accuracy, Yorùbá and European languagesdiacnet-2.0 →

      Lowest error on 8 of 10 languages (mean DER 0.0117). Meaning hints and optional typo correction.

    • Generative diacritics on a smaller budgetdiacnet-mini-2.0 →

      Same features at 300M, half the size, mean DER 0.0141.

    • Diacritics that can never change a letter, on CPUdiactag-2.0 →

      Character tagger: compliance 1.0000 by construction. 38.6 MB int8 ONNX, about 8 ms a sentence on a laptop CPU. Strongest on Igbo, Vietnamese and Modern Standard Arabic.

    / Run it your way

    Choose how to run it

    Olaverse API

    The fastest start: an API key, pay as you go, $1 of free credit.

    • ✓$1.00 per million characters with diacnet-2.0, $0.60 with diacnet-mini-2.0 and $0.20 with diactag-2.0.
    • ✓Calibrated outputs, no infrastructure
    • ✓Your data is never stored or used for training
    Get an API key

    Open weights

    Download the model and run it on your own GPUs.

    • ✓Hugging Face, open licence
    • ✓Keep every document in your environment
    • ✓Fine-tune on your own data
    Get the weights

    Enterprise

    A private deployment, tuned to your data and workflows.

    • ✓Private cloud or on-premises
    • ✓Fine-tuning on your labels
    • ✓Volume pricing and support
    Talk to us
    / FAQ

    Questions

    Will it change my words?+

    No. With alignment on (the default), every letter you typed is kept and only marks are added. diactag-2.0 goes further: it never predicts letters at all.

    Do I need to tell it the language?+

    No. Leave the language as auto and it detects it, or set it when you know.

    Which model should I use?+

    diacnet-2.0 for the best accuracy, diacnet-mini-2.0 for half the size, and diactag-2.0 when you need it on a CPU and a guarantee that letters never change.

    Is my text stored?+

    No. Text is processed in memory, never logged and never used for training.