Why the Voynich Manuscript Might Not Be a Real Language

[quimqu]

The Voynich Manuscript, transcribed using the EVA alphabet

Others have already pointed this out, but it feels worth repeating: this explains a significant part of the behaviour of the graph. 

Use modified transliteration alphabets and see what happens. 
At least convert Ch Sh in and iin to single characters.

Also a question: are space characters treated just like normal text characters?

Your result is quite different from my results here: You are not allowed to view links. Register or Login to view.
where Voynichese becomes less predictable after 3 characters, but that is probably due to the way in which I treated space characters, namely as separators.

Thanks — that’s a great point, and I actually followed your suggestion to test with a modified transliteration system.

I compared EVA and CUVA transliterations (following your 'rules file' conversion provided in the site You are not allowed to view links. Register or Login to view. ). This gives a more compact alphabet. (By the way, note that in EVA and CUVA I assigned each extended character glyph a distinct Unicode).



Interestingly, CUVA produces higher perplexity beyond n=3, which might seem counterintuitive. But here's what's happening:

• The CUVA glyphs increase entropy per character, because each unit contains more information.
  
• As a result, sequences become less predictable in higher-order n-grams — especially when you're conditioning on long contexts.
  
• So even though CUVA seems more compact, it compresses repeated structure into fewer units, reducing the surface-level redundancy that makes Voynichese so predictable under EVA.
  

This supports the idea that Voynichese's low perplexity (in EVA) reflects encoding regularity rather than linguistic simplicity.

Even in CUVA form, Voynichese is still more predictable than Latin or English — unless you use high-order n-grams, where CUVA becomes far less compressible.

Regarding Your Question About Spaces: Yes — in my case, spaces are treated as normal characters. I model character-level perplexity over the full stream, including word boundaries. This is crucial, since many of Voynichese’s patterns involve position within words (q-initial, -dy final, etc.). So the model learns transition probabilities across word endings too.

This may explain part of the divergence from your entropy graph, which treats space as a delimiter and resets statistics per word. That method is great for isolating internal word structure, while mine reflects stream-level predictability.

So in short:

• Glyph-grouping (like CUVA) affects predictability a lot — especially in high-order n-grams.
  
• EVA’s low perplexity seems to stem from highly redundant symbol sequences, not linguistic simplicity.
  
• Treating space as a token (vs delimiter) changes the entropy dynamics significantly.
  

Thanks again for raising this!