Voynich Alphabet Problem: How Many Characters Are There?

How many letters are in the Voynich Manuscript?

It sounds like the easiest question in the entire mystery.

Count the shapes.

Make an alphabet.

Begin deciphering.

Instead, the alphabet question is one of the deepest problems in Voynich research.

Jacques Guy summarised the difficulty bluntly decades ago: we cannot be certain what constitutes a Voynich letter.

Jim Reeds described the same problem from another direction. A highly analytical transcription can represent the manuscript with roughly twenty primitive symbols. A more synthetic system such as Currier’s can require around three dozen signs before rare characters are added. Once unusual one-off and compound forms enter the inventory, the apparent alphabet can grow dramatically.

The parchment did not change.

Only our answer to one question changed:

What counts as one character?

That decision propagates everywhere.

Character frequency.

Word length.

Entropy.

Word families.

Repeated strings.

Apparent morphology.

Even the plausibility of a cipher.

The alphabet therefore comes before decipherment in a stronger sense than most readers realise.

You cannot ask what the characters mean until you have a defensible theory of where one character ends and another begins.

Quick Read

One-sentence answer: there is no universally accepted single number of Voynich characters because different transcription systems place the character boundary at different levels—primitive strokes, likely graphemes, ligatures, compound gallows, repeated minim strings and rare forms—and the manuscript itself does not tell us directly which of those visible units functioned as independent signs.

The Voynich script clearly contains a stable core of frequent recurring shapes.
The exact size of that core depends on how compound-looking forms are segmented.
Currier’s transcription is more synthetic than EVA and treats some compounds as single units.
Frogguy and EVA are more analytical and expose recurring graphical components.
Basic EVA deliberately keeps a compact common character set.
Extended EVA adds rare and unusual forms so the visible manuscript can be represented more faithfully.
v101 is comparatively synthetic and gives many compound or ligature-like forms their own codes.
The modern Super Transliteration Alphabet (STA) was built as a superset capable of representing distinctions preserved by several major alphabets.
STA itself warns that one STA code should not automatically be interpreted as one medieval character; a code can represent a long ligature-like form.
Repeated i/e-like minims are a special problem because older alphabets sometimes treated strings such as i+n and ii+n as single synthetic characters.
Pedestalled gallows are another special problem because analytical systems split them while synthetic systems can treat them as units.
Rare “weirdo” signs can be genuine characters, compounds, scribal variants, mistakes, corrections or damaged ordinary signs.
Alphabet size therefore depends on palaeography, not only computation.
A good decipherment should make one character model explain the script more economically and predictively than its competitors.

So the right answer is not a magical integer.

It is a hierarchy of increasingly committed models.

The Ink Gives Us Shapes, Not an Alphabet Table

If the Voynich Manuscript contained a medieval primer that said:

these are the twenty-three letters of our script

the problem would be much easier.

Instead, we observe handwriting.

Handwriting is continuous.

Two strokes can connect.

One stroke can vary.

A flourish can be optional.

A familiar unit can be enlarged at a paragraph start.

A ligature can fuse two signs into one continuous gesture.

A compound character can be built from reusable components.

Every script asks palaeographers to distinguish:

character;
allograph;
ligature;
abbreviation;
compound;
decorative variant;
scribal error.

Voynich makes that ordinary palaeographic job harder because there is no securely identified external writing tradition against which the distinctions can be calibrated.

We are learning the alphabet and the variation model at the same time.

Analytical and Synthetic Alphabets

The central choice can be described with two words.

Analytical

Break complex visible forms into recurring parts.

If a sign visibly looks like frame + gallows + frame, represent those components separately.

Synthetic

Preserve complex visible forms as whole units.

If a writer appears to produce one integrated compound, give the compound one code.

Neither philosophy is automatically correct.

They answer different questions.

An analytical alphabet asks:

Which graphical pieces recur?
Can one gallows substitute inside the same frame?
Do repeated minims form longer structures?

A synthetic alphabet asks:

How often does this complete visible form occur?
Does the whole sign behave like one unit?
What happens if we do not manufacture internal transitions through decomposition?

The Voynich alphabet problem lives on a continuum between the two.

Why One Manuscript Can Have Twenty, Thirty-Six or Far More “Characters”

Jim Reeds once framed the problem with a useful contrast.

A highly analytical Frogguy representation could work with roughly twenty primitive symbols.

Currier’s system used thirty-six.

Then rare signs discovered across the manuscript could add dozens more apparent forms.

These numbers are not contradictions.

They describe different levels.

Imagine English handwriting.

At one level we say it has twenty-six letters.

At another level we distinguish uppercase and lowercase.

At another, ligatures such as æ or historical scribal abbreviations receive separate treatment.

At another, every handwritten allograph could be considered a distinct visual form.

The number grows because the ontology changes.

Voynich suffers the same problem without an external grammar telling us which level is linguistically privileged.

Basic EVA Is Deliberately Not the Whole Manuscript

EVA became popular because it is practical.

Its basic character set captures the common writing with a relatively compact inventory and makes electronic comparison straightforward.

But the existence of basic EVA should never be mistaken for proof that the manuscript has exactly that many underlying graphemes.

Extended EVA exists precisely because the manuscript contains rare and unusual forms that basic EVA cannot represent faithfully without extra codes or annotations.

The latest definitions also preserve ligatures explicitly.

This creates a useful distinction:

basic alphabet for common analysis ≠ complete palaeographic inventory.

One is a working abstraction.

The other tries to preserve exceptional ink.

Both are needed.

Extended EVA Creates a Long Tail

Once rare characters are admitted, alphabet counting becomes unstable quickly.

One odd form may appear only once.

Another may be a common sign with an added flourish.

Another may be damage.

Another may be a true compound.

Another may be a later mark.

Extended alphabets must record the distinction before they know what it means.

This is scientifically healthy.

It avoids destroying evidence by normalising every rare form into a common character.

But it means the number of transliteration codes cannot be equated directly with the number of functional graphemes.

An extended alphabet is an observational inventory.

A functional alphabet is a theory.

v101 Shows the Other Philosophy

The v101 transliteration tradition is more synthetic.

Combined and ligature-like forms are more often represented as units rather than decomposed into EVA-style character strings.

This can make the alphabet look larger.

But the sequence can become less artificially predictable because internal transitions inside a compound are no longer counted as sequential characters.

That trade-off matters for entropy.

If one visual glyph is written in EVA as three characters, the first strongly predicts the second and the second the third.

Conditional entropy falls.

If the same visible glyph receives one synthetic code, those internal transitions disappear.

The manuscript did not become less regular.

The representation stopped measuring one kind of internal structure as character sequence.

This is why alphabet size and entropy cannot be discussed separately.

STA Was Built Because No Earlier Alphabet Preserved Everything

The Super Transliteration Alphabet is a modern attempt to solve a practical interoperability problem.

Different major transliterations preserved different distinctions.

Some forms existed in extended EVA but not v101.

Others existed in v101 but not extended EVA.

STA therefore acts as a superset capable of aligning those representations.

This is a major infrastructural improvement.

It is not an announcement that every STA code is one true Voynich letter.

The STA documentation says the opposite explicitly: because the alphabet adopts a synthetic strategy in places, one code can represent a long ligature-like form and may correspond to more than one underlying character.

This is exactly the right epistemic separation.

A transliteration code identifies a visible analytical unit. It does not automatically identify the manuscript’s linguistic atom.

Pedestalled Gallows Can Inflate or Deflate the Alphabet

The gallows and bench articles already established the visible architecture.

Take a pedestalled form written analytically as c-k-h.

One model counts three components.

Another counts one compound grapheme.

A third counts a frame plus a payload: two functional units.

Three alphabet sizes emerge from one ink shape.

Which one is correct?

The strongest answer should come from several independent tests.

Stroke construction.
Substitutability.
Distribution.
Entropy under alternative segmentations.
Decipherment performance.

If the frame contributes a stable function and the inserted gallows contributes another, a compositional alphabet wins.

If the complete sign behaves atomically, a synthetic grapheme model wins.

The alphabet should be inferred from behaviour, not visual intuition alone.

The Minim Strings Create a Second Alphabet Problem

Repeated i-like strokes are another classic trouble spot.

Older alphabets sometimes treated i+n and ii+n as synthetic characters.

EVA exposes the repeated i-like components.

STA deliberately treats strings of i and c as a special problem because no representation is obviously privileged.

Are i, ii and iii repetitions of one grapheme?

Are i+n and ii+n equivalent to synthetic forms historically transcribed as N and M?

Is final n simply a contextual final form of i?

Does the number of minims encode quantity?

These alternatives change the alphabet drastically.

The next article owns the minim system in depth.

Rare Glyphs Create a Third Alphabet Problem

Jim Reeds’ “weirdoes” made a simple alphabet impossible.

Some unusual signs were not represented in older standard inventories.

Jacques Guy expanded Frogguy precisely so these forms could be recorded.

But recording a rare form does not prove it is a new letter.

One rare sign may be:

a true low-frequency grapheme;
a compound of common elements;
an allograph;
a scribal mistake;
a damaged form;
a correction;
a later addition.

If every visual novelty becomes a new character, the alphabet explodes.

If every novelty is normalised away, real distinctions disappear.

The third article in this batch owns that rare-glyph long tail.

Allographs Can Make One Character Look Like Several

Every handwriting system contains variation.

The same writer forms one letter slightly differently at different times.

Different scribes exaggerate the variation.

A formal opening may receive a larger form.

A line ending may receive a flourish.

Palaeography calls functionally equivalent variants allographs.

If Voynich contains several hands—or even one hand changing through time—some apparent alphabet expansion may be allographic.

This is why Lisa Fagin Davis’s palaeographic work matters to alphabet counting as well as scribe counting.

Glyph classification and hand classification interact.

A form unique to one proposed scribe may be a personal variant rather than a separate character.

Alphabet size cannot be settled without a model of handwriting variation.

Majuscules and Positional Variants Could Inflate the Count Too

Some scripts distinguish forms by position or status.

Uppercase and lowercase are the familiar example.

Arabic has contextual letter forms.

Historical scripts contain final forms and display variants.

Voynich has strong line and paragraph positional effects.

Therefore some visible variants may be contextual rather than lexical.

If a line-final form is merely the final allograph of an ordinary sign, treating it as a separate character inflates the functional alphabet.

If it carries a different grammatical or encoded value, collapsing it loses information.

Again, position gives us a discriminator.

Contextual forms should be predictable from context.

Independent characters should carry information beyond that context.

An Alphabet Is Not Necessarily Phonetic

There is another hidden assumption in the phrase “Voynich alphabet”.

Alphabet suggests sounds.

But the script could encode:

phonemes;
syllables;
morphemes;
abbreviations;
code groups;
operators;
mixed units.

A system with twenty visible primitive components may encode hundreds of linguistic values through compounds.

A system with one hundred synthetic signs may reduce to a much smaller set of functional features.

Therefore character count alone says little about language type.

Chinese characters, syllabaries, alphabets, abjads and cipher tables all map visible units to information differently.

Voynich needs its own mapping recovered rather than assumed from the word “alphabet”.

Alphabet Size Changes Entropy

The entropy article already established that Voynichese is unusually predictable at character level.

But entropy depends on the symbols being counted.

Split one complex sign into three components.

You create two internal transitions.

If those components almost always occur together, the transitions are highly predictable.

Merge them into one symbol.

The transitions disappear, but the alphabet expands.

This is not statistical trickery.

It is the unavoidable consequence of representation.

Lindemann and Bowern therefore explored glyph-composition alternatives rather than assuming EVA segmentation was sacred.

The broad unusual predictability persisted under substantial reasonable manipulations.

That strengthens the phenomenon while keeping character ontology open.

Alphabet Size Changes Word Length

A word written with five EVA characters may contain three synthetic graphemes under another model.

This matters because researchers compare Voynich word lengths with known languages.

If the grapheme count is wrong, the comparison is answering the wrong question.

Likewise, abbreviation can make a short visual token represent a long underlying word.

Word length therefore depends on at least three layers:

Where visible spaces divide tokens.
How visible compounds are divided into graphemes.
How graphemes map to underlying linguistic material.

A numerical average at Layer 1 cannot be compared naively with a linguistic average at Layer 3.

Alphabet Size Changes Word Families

Edit distance depends on the alphabet too.

If ckh is three characters, replacing k with t is one edit.

If ckh and cth are atomic graphemes, replacing one with the other is still one edit—but the internal relationship between k and t disappears.

If the frame and payload are independent features, the most faithful distance measure may not be ordinary character edit distance at all.

The same issue occurs with minim strings.

iiin and iin may differ by one minim under an analytical model but be two unrelated synthetic characters under another.

Word-family geometry therefore depends on character ontology.

A robust family should survive reasonable representations or become more interpretable under one specific model.

Alphabet Size Changes Cipher Plausibility

A simple monoalphabetic substitution with one hundred independent symbols is a different object from a twenty-symbol system with productive ligatures.

The key space differs.

Frequency analysis differs.

Historical plausibility differs.

A verbose cipher may deliberately map one plaintext element to a multi-sign group.

An abbreviation system may compress several plaintext letters into one sign.

The same visible manuscript can therefore look like an expanded cipher under one segmentation and compressed shorthand under another.

This is why claims such as “the alphabet is too small” or “the alphabet is too large” are not decisive without an explicit unit model.

Palaeography Must Arbitrate the Statistical Models

Computation can tell us that one segmentation compresses the corpus efficiently.

It can tell us that another improves language-model fit.

It can show that a third reduces conditional entropy anomalies.

But a segmentation that contradicts how the signs are physically written is suspect.

Palaeographic evidence includes:

stroke continuity;
pen lifts;
connection patterns;
variation across hands;
positional variants;
recurring construction order.

The best alphabet model should make both the ink and the statistics simpler.

If computational elegance requires treating one obvious continuous ligature as three independent letters, something is wrong.

If palaeographic intuition creates hundreds of atomic characters with no recurring structure, something may also be wrong.

The correct ontology should reconcile both.

A Functional Alphabet May Be Smaller Than the Visible Inventory

Suppose the manuscript has forty common visible forms.

Twenty of them are combinations of ten reusable components.

The visible inventory is forty.

The functional component inventory may be much smaller.

Now suppose several components are merely contextual allographs.

The underlying inventory shrinks again.

This layered view may be the most realistic way to think about Voynich.

Stroke primitives.
Grapheme components.
Functional graphemes.
Ligatures and compounds.
Rare palaeographic forms.

There may be no single alphabet number that is useful at every layer.

The research job is to recover the hierarchy.

The Correct Alphabet Should Make Rare Forms Less Arbitrary

A good model explains the long tail.

If rare signs can be generated as predictable combinations of common components, the effective alphabet shrinks.

If some rare forms are hand-specific allographs, the functional alphabet shrinks.

If a few rare forms have independent distributions that no component model predicts, the alphabet may genuinely need to expand.

The rare signs are therefore not noise around the alphabet.

They are a hostile test of the alphabet model.

A theory that explains only the twenty commonest shapes is incomplete.

A theory that invents a separate meaning for every rare flourish is overfit.

The right model should compress rarity without erasing genuine exceptions.

The Correct Alphabet Should Predict Unseen Writing

Imagine a new folio fragment were discovered tomorrow.

A strong alphabet model should tell us:

which new visible forms are legal compounds;
which would count as new graphemes;
which would likely be allographs;
how frequent combinations should be segmented;
which rare forms would genuinely expand the inventory.

This is a better standard than asking whether the model can explain every form after seeing it.

A character theory should be generative.

It should tell us what kinds of written objects the script is capable of making.

What Survives the Alphabet Work

The Voynich script has a stable core of recurring common forms.
No universally accepted single functional alphabet size exists.
Analytical and synthetic transcription philosophies produce different inventories from the same ink.
Basic EVA is a practical common-character abstraction, not proof of the underlying alphabet size.
Extended EVA preserves rare forms and ligatures that the basic set omits.
v101 and STA preserve more synthetic compound distinctions in important places.
STA is a superset/interoperability alphabet, not a declaration that every code equals one linguistic character.
Pedestalled gallows, bench forms, minim strings and rare glyphs are the main character-boundary stress tests.
Scribal allographs and positional variants can inflate visible inventory.
Character segmentation changes entropy, word length, word families and cipher interpretation.
Palaeographic and statistical evidence must converge before a functional alphabet can be considered strong.

What Does Not Survive as Established Knowledge

The Voynich Manuscript has exactly the number of characters in basic EVA.
Every extended-EVA code is an independent grapheme.
Every STA code is one medieval letter.
Every visible ligature is one grapheme.
Every visible component is an independent grapheme.
Rare signs automatically expand the alphabet.
Rare signs can safely be normalised away.
The alphabet size by itself identifies a language family.
The alphabet size by itself proves or disproves a cipher.

The visible diversity survives.

The single definitive integer does not.

A Better Alphabet Analysis

Keep visible form and functional grapheme separate.
Preserve rare forms before normalising them.
Compare analytical and synthetic alphabets.
Track ligatures explicitly.
Test gallows-plus-bench composition.
Treat minim strings as a special subsystem.
Model allographs by hand and position.
Measure how each segmentation changes entropy and word families.
Require the alphabet to improve predictive decipherment, not just descriptive neatness.
Keep every abstract code traceable to the manuscript image.

What Would Count as a Real Alphabet Breakthrough?

Imagine several independent palaeographers classify the script from high-resolution images without using one existing transliteration as the answer key.

They converge on a compact set of functional graphemes plus a small set of predictable ligature and positional rules.

The same model explains pedestalled gallows, minim strings and most rare forms.

When the corpus is retranscribed under that model:

entropy becomes more historically plausible;
word families become cleaner;
class-level syntax strengthens;
a decipherment succeeds on unseen pages.

That would be a true alphabet breakthrough.

Or the opposite might happen.

A synthetic inventory of many compounds may consistently outperform decomposition and prove that Voynich uses a larger set of atomic written units than expected.

The important point is not whether the final number is small or large.

The correct alphabet is the one that makes the rest of the manuscript easier to explain without inventing exceptions.

Primary School: How Many Symbols Are in This Symbol?

Draw a circle with a star inside.

Ask a child how many symbols they see.

One answer: one compound symbol.

Another: a circle plus a star.

Neither answer is silly.

The correct answer depends on how the writing system uses the parts.

This is the Voynich alphabet problem in its simplest form.

Lower Secondary: Two Alphabets From One Dataset

Give students ten compound symbols made from four reusable parts.

Group A assigns one code to each complete symbol.

Group B writes each symbol as a sequence of parts.

Ask:

Which alphabet is larger?
Which text is longer?
Which has more predictable adjacent symbols?

The students discover that representation changes statistics before any meaning changes.

Upper Secondary: Allograph or New Letter?

Show several handwritten versions of one familiar letter from different people.

Then include one genuinely different letter that looks similar.

Ask students to decide where variation becomes category change.

They will discover that there is no purely visual threshold.

Distribution and function matter.

That is exactly the challenge posed by Voynich rare glyphs and scribal variation.

JC and Adult Readers: Alphabet Recovery Is Model Selection

At a higher level, the alphabet problem can be formalised as competing latent ontologies.

Model A has a small primitive alphabet and productive composition.

Model B has a larger atomic grapheme inventory.

Model C adds contextual allographs.

Model D includes abbreviation expansions.

The correct comparison is not which model fits the observed corpus best with unlimited parameters.

It is which model best predicts held-out forms, reduces arbitrary exceptions, agrees with palaeography and supports independent decipherment.

Alphabet recovery is therefore an inference problem, not a counting exercise.

A Parent and Teacher Guide

Ask what counts as one unit before counting.
Separate visible form from functional character.
Compare analytical and synthetic representations.
Keep rare forms instead of erasing them early.
Remember that handwriting creates allographs.
Show how alphabet choice changes entropy and word length.
Require a final alphabet to improve prediction, not merely tidy the data.

The wider reasoning lesson is simple:

before counting categories, first prove where the category boundaries belong.

Reader Checklist: Before You State the Size of the Voynich Alphabet

Which transliteration alphabet are you using?
Is it analytical or synthetic?
Are pedestalled gallows one unit or several?
How are ch/sh benches represented?
How are i/ii/iii minim strings represented?
Are rare glyphs included?
Are allographs separated by hand?
Are positional forms separate codes?
Are later or doubtful marks excluded?
How does the alphabet choice change entropy?
How does it change word length?
Does the model predict rare and unseen compounds?
Does palaeography support the segmentation?
Does decipherment improve under the same model?

Frequently Asked Questions

How many characters are in the Voynich alphabet?

There is no single universally accepted number. Compact analytical alphabets can represent the writing with a relatively small set of primitive elements, while synthetic and extended inventories preserve many more compound and rare forms.

Why can different researchers get different numbers?

Because they disagree about whether compound forms, ligatures, minim strings, positional variants and rare signs should be counted as separate characters or decomposed into reusable parts.

Is EVA the real Voynich alphabet?

No. EVA is a highly useful transliteration system. Its Roman letters are analytical labels, and its segmentation choices are not decoded facts about the medieval writing system.

What is extended EVA?

It expands the basic EVA inventory to represent rare, unusual and ligatured forms more faithfully.

What is STA?

The Super Transliteration Alphabet is a modern superset designed to align and preserve distinctions found across major Voynich transliteration systems. Its codes are not automatically one-to-one with functional medieval letters.

Why do rare glyphs matter?

They test whether the alphabet model can explain unusual forms as compounds or allographs, or whether genuine low-frequency graphemes must be added.

Why does alphabet size matter for entropy?

Splitting one compound into several characters creates additional, often highly predictable transitions. Merging the compound removes those transitions and enlarges the alphabet.

Could the functional alphabet be very small?

Possibly, if many visible forms are built compositionally from a small set of primitives. It could also be larger if many compounds function atomically. The evidence has not settled the issue.

What is the strongest current conclusion?

The Voynich script has a stable core but an unresolved character ontology. A serious solution must explain which visible forms are primitives, compounds, ligatures, allographs and rare genuine graphemes.

Research and Further Reading

The Final Idea

The alphabet problem looks like bookkeeping until you realise that every later conclusion inherits it.

Count a bench as two characters and one entropy profile appears.

Count it as one and another appears.

Treat iiin as four signs and one morphology appears.

Treat it as one synthetic unit and another appears.

Normalise a rare flourish and the alphabet shrinks.

Preserve it and the alphabet grows.

The manuscript has not moved.

Our ontology has.

That is why the final solution will need more than a translation table.

It will need a theory of the script itself.

Before Voynich can tell us what its characters say, it still has to teach us what its characters are.

Continue Through the Voynich Research Map

This article is one specialist node in eduKateSG’s larger Voynich research library. Return to the canonical master to see the physical, visual, textual and historical evidence in one continuous argument.

Structure is evidence; structure is not translation. The master keeps resemblance, statistical structure, historical possibility and decipherment claims at separate evidentiary levels.

Next Best Routes

EVA, Transcription and the Segmentation Problem — return to the corpus representation that makes alphabet counting possible.
Rare Glyphs — use the long tail as a hostile test of any proposed inventory.
The Minim Strings — inspect one of the hardest cases for deciding whether repetition means several characters or one synthetic unit.

Deep bridge: The Model Is Part of the Message frames alphabet recovery as model selection: smaller inventories need stronger composition rules; larger inventories pay more in symbol complexity.

Voynich | Everything eduKate Knows and Tested | The Alphabet Problem: How Many Voynich Characters Are There?