Voynich Language Identification: Can We Know the Language Before Reading It?

Can we identify a language before we can read it?

Sometimes.

A known script may preserve letter frequencies.

A known language family may preserve characteristic sound patterns.

Word endings may expose morphology.

Function-word classes may reveal grammar.

Even an undeciphered inscription can sometimes be placed near a language family through repeated names, structural parallels or bilingual anchors.

Voynich is harder.

We do not know what one visible glyph represents.

We do not know whether a space is always a word boundary.

We do not know whether vowels are explicit.

We do not know whether the system is alphabetic.

We do not know whether it is shorthand.

We do not know whether it is ciphered.

We do not know whether every visible token maps to speech at all.

That means source-language identification is downstream of several unresolved representation problems.

And yet researchers keep trying.

Latin.

Italian.

Germanic languages.

Greek.

Hebrew.

Arabic or other Semitic systems.

Romance varieties.

English.

The difficulty is not lack of candidates.

It is that many representation choices can make one candidate look statistically attractive.

The Language-Identification Problem is not which language can be made to resemble Voynich. It is which source-language family remains the best explanation after the writing system, segmentation and transformation rules are frozen.

Quick Read

Direct answer: no underlying Voynich language has been identified by scholarly consensus. Linguistic research supports treating natural language as a serious hypothesis and shows that phonological, morphological and structural arguments can be made before decipherment. But language attribution remains highly sensitive to character definition, word boundaries, vowel representation and any cipher, shorthand or notation layer. Current 2026 claims point in mutually incompatible directions, which makes independent out-of-sample prediction more important than ever.

EVA is a transliteration convention, not a pronunciation.
Visible letter frequency cannot be compared directly with Latin or Italian unless Voynich glyphs are known to represent comparable units.
Word-length distributions change when compound glyphs or uncertain spaces are redefined.
Vowel omission can make unrelated languages look more similar.
Homophonic cipher can flatten language-specific frequencies.
Verbose cipher can alter word lengths and token families.
Shorthand can collapse several sounds into one visible sign.
Technical notation can remove source-language signals almost entirely.
Currier A/B means the manuscript should not be averaged into one homogeneous corpus automatically.
Labels, prose and record-like text may have different language statistics even under one source language.
Language identification must control genre as well as script.
Bowern and Lindemann argue that natural-language analysis remains productive even while content is undeciphered.
A 2026 Greek project claims an EVA-to-Greek mapping and remains under peer review.
A 2026 Italian shorthand preprint proposes syllabic tachygraphy rather than direct alphabetic Italian.
An August 2026 Zenodo preprint claims a 1,609-line English translation through procedural grammar while explicitly not recovering historical phonology.
Mutually incompatible success claims are evidence that evaluation standards—not candidate languages—are the bottleneck.
No candidate has yet produced accepted manuscript-wide language identification under stable independently validated rules.

Language Identification Is Not Translation

It is possible to know a language family without understanding every word.

Historical linguists routinely classify languages from sound correspondences, morphology and grammar.

Cryptanalysts can sometimes detect plaintext language statistically before a key is fully recovered.

But Voynich adds uncertainty one layer earlier.

What are the comparable units?

If we get that wrong, language statistics can become statistics of our transcription rather than of the historical language.

EVA Letters Are Not Sounds

EVA was designed to let researchers write and compare Voynich glyphs consistently.

It deliberately uses convenient Latin letters.

EVA q is called q because the glyph resembles q.

That does not mean /k/.

EVA o is not known to be /o/.

EVA ch is not known to be the sound written “ch” in English, Italian or German.

Any language-identification method that reads EVA letter names phonetically has already assumed part of the solution.

Character Inventory Comes Before Language Inventory

How many functional characters does Voynich have?

That remains unsettled.

Are bench forms single units?

Are gallows modified compounds?

Are rare signs independent letters?

Are minim strings one unit or several?

Different answers change frequency, entropy and apparent phonotactics.

The Alphabet and Unit-Scale problems therefore sit upstream of language identification.

Word Boundaries Come Before Word Statistics

If Voynich spaces correspond perfectly to lexical words, whole-token comparison is natural.

If some spaces divide larger units, token vocabulary is inflated.

If some missing spaces join words, vocabulary changes again.

Recent representation-focused work has made this old problem harder to ignore.

A candidate language should survive reasonable alternative space treatments rather than require one convenient segmentation.

The Vowel Problem Distorts Language Family Signals

Vowels carry strong language-specific information.

Remove them and distributions converge.

Arabic, Hebrew and heavily abbreviated Latin can all look more consonantal than ordinary modern European spelling.

A syllabic shorthand can hide vowels inside signs.

A cipher can redistribute them.

Therefore language identification from visible vowel-like glyph frequencies is premature until the representation is understood.

Phonotactics Is Stronger Than Letter Resemblance

A good language candidate should eventually explain legal sound sequences.

What can begin a syllable?

Where can vowels occur?

Which consonant clusters are possible?

The Phonotactics Problem owns that downstream system.

Language identification should use it as validation.

A Latin claim that requires Latin words to violate one supposedly recovered Voynich phonology is internally inconsistent.

Morphology Should Travel Beyond One Word

Latin is highly inflected.

Italian is less so but still morphological.

Greek has characteristic endings and agreement patterns.

Semitic languages have root-and-pattern structures of another kind.

A genuine language identification should recover family-appropriate morphology across many lexical roots.

One Voynich ending cannot become a Latin case marker in one word and an unrelated syllable in the next whenever convenient.

Function Words Are Another Language-Family Test

Languages organise grammatical relations differently.

Latin lacks an English-style article system.

Greek has articles.

Romance languages use characteristic prepositions and clitics.

A candidate language should therefore predict a plausible closed grammatical class after representation is recovered.

The Function-Word Problem gives that claim a structure-first test before dictionary meanings are assigned.

Currier A and B Can Fake Two Languages—or Reveal Them

Currier famously labelled two strong textual regimes A and B.

They could be:

two languages;
two dialects;
two orthographies;
two cipher states;
two shorthand conventions;
two document roles;
two stages of one evolving system.

A language-identification study that mixes A and B into one corpus can blur real signals.

A study that separates them automatically can manufacture two source languages from one representation switch.

The Currier Switch and Drift problems therefore become mandatory controls.

Genre Can Mimic Language Difference

A recipe book and an astronomical table written in the same language have different vocabularies.

A label set has different word lengths from prose.

A glossary differs from narrative.

The newly published Genre Problem therefore matters directly.

Language models must compare matched document roles rather than attribute every vocabulary shift to dialect or language family.

Latin Is Historically Attractive

Latin was a major learned language in fifteenth-century Europe.

Medicine, astrology, natural philosophy and manuscript production all make it an obvious candidate.

Medieval Latin also used heavy abbreviation.

That can alter visible word lengths and vowel patterns.

But historical plausibility is not statistical identification.

A Latin solution must recover stable Latin morphology and syntax under one mapping.

Italian Is Historically Attractive Too

Northern Italian manuscript cultures sit near several long-running Voynich provenance hypotheses.

Vernacular medical writing existed.

Notarial shorthand existed.

The 2026 shorthand preprint specifically proposes Italian syllabic tachygraphy.

Again, a plausible historical environment is not a recovered source language.

The model must expand unseen text into connected period-appropriate language.

Germanic Candidates Have Historical Reasons Too

Later provenance connects the manuscript to Central European worlds.

Some Voynich glyph shapes and marginalia have encouraged German-related hypotheses.

But later ownership does not establish original language.

A Germanic candidate has to win linguistically rather than geographically.

Greek Has a New 2026 Claim

Christos Papanikolaou’s Nereus Project released a 2026 reproducibility package for a proposed EVA-to-Greek mapping.

The associated paper is described as under peer review and claims a statistical identification of a Greek shorthand writing system.

The package is valuable because it exposes rules, corpus and scripts rather than offering only prose interpretation.

Its existence does not make Greek established.

Independent researchers still need to verify:

mapping uniqueness;
dictionary-collision rates;
historical shorthand ancestry;
morphological coherence;
held-out decoding;
Currier transfer.

Semitic Candidates Remain Mechanistically Interesting

Hebrew and Arabic-related hypotheses recur partly because omitted-vowel writing can resemble some Voynich statistical properties.

This is a mechanism-level reason to test them.

It is not evidence that Voynich is Semitic.

A real identification should recover root-and-pattern morphology or other family-specific structures under one fixed representation.

English Now Has an August 2026 Claim Too

An August 2026 Zenodo preprint by Matthew Dominik presents a 1,609-line English translation derived from a frozen procedural grammar.

The deposit reports high structural-role coverage on a constrained benchmark and includes reproducibility materials.

It explicitly states that it does not recover historical phonology or a complete lexical dictionary.

This makes the claim methodologically unusual.

It proposes English semantic output without a traditional sound-by-sound decipherment.

That is precisely why it should not be used as evidence that “English has now been identified”.

The claim requires independent scrutiny of grammar assignment, lexical role mapping, semantic constraints and benchmark construction.

Mutually Incompatible Successes Are a Diagnostic Signal

Greek cannot simultaneously be the unique underlying language if another equally flexible system uniquely proves English and a third proves Italian shorthand—unless these claims operate at genuinely different layers.

When incompatible theories all report strong fits, several possibilities arise.

one is right and others overfit;
all are overfitting;
the evaluation metrics are too weak;
the manuscript representation allows many mappings;
different projects are identifying different structural layers rather than the same language.

This is why the Joint Benchmark and Model Complexity problems matter.

The more candidate languages can pass the same test, the less that test identifies language.

Language-Like Structure Does Not Identify a Language

Bowern and Lindemann’s Annual Review argues that there are good reasons to treat natural language as a serious model and demonstrates how linguistic analysis can proceed before decipherment.

That is important.

But “language-like” is a category-level conclusion.

It does not choose Latin over Greek.

Or Italian over German.

Specific identification requires finer discriminators.

Language Identification Must Survive Cipher Controls

The 2025 Naibbe cipher is a major warning.

Meaningful Latin or Italian passed through a verbose homophonic cipher can acquire several Voynich-like statistics.

Therefore visible Voynich statistics may identify the encryption mechanism more strongly than the plaintext language.

A good source-language classifier should be tested on encrypted controls.

Can it still distinguish Latin from Italian after the proposed transformation?

If not, it cannot identify Voynich source language under that mechanism.

Language Identification Must Survive Shorthand Controls

The same applies to shorthand.

Compress Latin and Italian through a syllabic notation.

How much family signal survives?

Which features remain diagnostic?

A classifier trained on ordinary spelling may fail completely.

Language Identification Must Survive the Notation Alternative

The newly published Notation Problem adds an even deeper challenge.

If visible Voynich units represent functions or classes rather than speech, source-language identification at the visible layer may be meaningless.

The right question would become:

what linguistic environment, if any, generated or accompanied the notation?

That may require external historical evidence rather than token statistics alone.

A Real Language Candidate Should Compress the Unknowns

This is the most useful practical rule.

Start with a candidate language.

Recover a mapping on one bounded corpus.

Freeze it.

Move to unseen text.

If the candidate is correct:

new phoneme assignments should decline;
new grammar exceptions should decline;
morphology should recur;
function words should stabilise;
proper names should become constrained;
source spelling should become historically coherent.

The model should become easier to use as the corpus grows.

If every folio needs new values, the “language” is merely following the text.

Proper Names Are the Late-Stage Test

Proper names are attractive early cribs and dangerous ones.

Too many plant, star and place candidates exist.

A mature language model should reach names late.

After phonology.

After morphology.

After grammar.

Then labels become powerful independent checks.

The Best Language Test Is Prediction, Not Resemblance

Resemblance says:

this Voynich form looks like a Greek word.

Prediction says:

under a rule frozen before this folio was examined, this unknown form must take one of three Greek endings, and only one occurs.

The second is much stronger.

What Survives the Language-Identification Work

No underlying Voynich language has been accepted by scholarly consensus.
Natural language remains a serious hypothesis supported by substantial linguistic structure.
Language-family analysis can be meaningful before full translation when units are reliable.
Voynich’s functional units and word boundaries remain uncertain.
Vowel representation is unresolved.
Currier, genre and document role confound simple corpus-wide language classification.
Cipher, shorthand and notation can destroy or transform language-specific surface signals.
Latin, Italian, Greek, Germanic and Semitic candidates remain hypotheses rather than established identifications.
2026 has produced incompatible new language/translation claims, increasing the importance of common benchmarks.
A real identification should make the mapping and grammar converge on unseen text.

What Does Not Survive as Established Knowledge

Voynich is proven Latin.
Voynich is proven Italian.
Voynich is proven Greek.
Voynich is proven Hebrew or another Semitic language.
Voynich is proven English.
EVA letters can be compared directly with ordinary alphabet frequencies as phonemes.
One high-scoring language classifier proves source language.
A handful of dictionary matches establishes identification.
Geographic plausibility establishes language.
One 2026 preprint has changed the consensus status to deciphered.

A Better Language-Identification Analysis

Freeze a functional character representation.
Test alternate reasonable segmentations.
Separate Currier/RZ states and document roles.
Test genre-matched corpora.
Compare candidate languages under matched orthographic transformations.
Include vowel-omission, shorthand and cipher controls.
Penalise mapping multiplicity and exception rules.
Learn on one bounded corpus.
Freeze values and grammar.
Predict entire unseen bifolia or folios.
Require morphology, function words and phonotactics to agree.
Use proper-name labels only as late independent validation.

What Would Count as a Real Language-Identification Breakthrough?

Imagine one candidate language wins under several independent representations.

The classifier is trained on transformed historical controls appropriate to the proposed shorthand or cipher mechanism.

It still selects the same language family.

A compact mapping is frozen.

On unseen bifolia, source morphology appears with the expected historical endings.

Function words occupy stable grammatical roles.

Phonotactics become legal under one sound system.

Currier A/B resolve into one explainable dialect, orthographic or cipher relationship.

Labels recover plausible proper names without new mapping rules.

Independent teams reproduce the result.

At that point, language identification would precede full translation legitimately.

The underlying language will be identified when it stops being the language that can explain Voynich and becomes the language that Voynich will no longer let us avoid.

Primary School: Same Secret, Different Language

Write the same simple substitution cipher using an English sentence and a Chinese sentence transliterated into letters.

The visible symbols hide both.

A child sees that recognising the cipher symbols is not the same as knowing the source language.

Secondary School: Remove the Vowels

Take short samples from several languages.

Remove vowels and spaces.

Compare again.

Many easy language clues disappear.

This demonstrates why representation changes the identification problem.

JC and Adult Readers: Source Attribution Under an Unknown Channel

Formally, Voynich language identification is source attribution through an unknown channel.

The source may be Latin, Italian, Greek or another language.

The channel may omit vowels, merge phonemes, apply shorthand, encrypt, transpose or encode functions.

Observed Voynichese is the output.

The identification problem is not solvable robustly until source and channel are inferred jointly or one is independently constrained.

Reader Checklist: Before You Believe a Voynich Language Identification

What are the functional Voynich units?
Are EVA letters being treated as sounds?
How are uncertain spaces handled?
Where are vowels?
What transformation channel is assumed?
Are Currier A/B separated or modelled?
Are labels mixed with prose?
Are controls matched by genre?
How many candidate languages were tested?
How many free mappings are allowed?
Was the mapping frozen before held-out text?
Does morphology emerge?
Do function words emerge?
Does phonotactics emerge?
Can an independent team reproduce the result?

Frequently Asked Questions

Do we know what language the Voynich Manuscript is written in?

No. The underlying language, if ordinary natural language is present, remains unidentified by scholarly consensus.

Can statistics identify it?

Potentially, but only after controlling representation, segmentation, genre and possible transformation mechanisms. Surface statistics alone are not sufficient.

Has Greek been proven in 2026?

No. A reproducible Greek-mapping project has been released and its associated paper is described as under peer review, but Greek is not an accepted identification.

Has the 2026 English translation solved Voynich?

No scholarly consensus has adopted it. The Zenodo preprint reports a procedural translation benchmark and explicitly does not recover historical phonology or a complete lexical dictionary.

What is the strongest current conclusion?

Language analysis is productive and natural language remains a serious model, but no candidate language yet survives the full representation, mechanism and held-out-prediction burden.

Research and Further Reading

The Final Idea

The Voynich Manuscript has never lacked languages.

Researchers have offered them for generations.

What it lacks is one language that keeps winning after the rules are frozen.

A candidate must survive the glyph problem.

The space problem.

The vowel problem.

Currier.

Genre.

Shorthand.

Cipher.

Notation.

And then it must still predict language on pages it has never seen.

The language underneath Voynich will become convincing when every new folio reduces our freedom to choose another language rather than increasing it.

Voynich | Everything eduKate Knows and Tested | The Language-Identification Problem: Can We Identify the Language Before We Can Read It?