VOICEPRINT — Where Am I in This Flow?

01

Origin

How It Began

I did not begin with books.

When I was young, I simply wrote down thoughts. Sometimes it was one sentence. Sometimes a few lines. Then a few pages. Then notebooks began to appear.

Gradually, those notes started connecting with one another. One thought continued another, and short fragments grew into longer texts.

At some point, I first thought: why not gather all of this into a book?

But then it turned out that there were walls between what had been written and the reader.

Read the full story ↓

The Full Story

I did not begin with books.

When I was young, I simply wrote down thoughts. Sometimes it was one sentence. Sometimes a few lines. Then a few pages. Then notebooks began to appear.

I wrote down observations, questions, doubts, individual images, certain conclusions — everything I did not want to lose.

Gradually, those notes started connecting with one another. One thought continued another, and short fragments grew into longer texts.

At some point, I first thought: why not gather all of this into a book?

That was when I discovered that writing a book and being able to bring it to a reader are not the same thing at all.

In the country where I lived at the time, I could not freely publish much of what I wrote about — or the way I looked at the human being, society, life, and reality. It was not simply a matter of a publisher possibly rejecting a manuscript. Publicly expressing certain views could itself have serious consequences.

So the most obvious path — write a book and publish it there — was effectively closed to me.

There was another path: to move beyond a single language.

But that brought a different problem.

I did not know foreign languages well enough to carry a complex literary and philosophical text into them myself. And professional translation of large works cost so much that, for me, it was practically inaccessible.

So texts could exist for years — in notebooks, files, drafts. Not because I did not want to bring them to readers, but because I could not see a realistic way to do it.

With the arrival of modern AI systems, the situation began to change for the first time.

It became possible to work with large volumes of text, bring scattered materials together, maintain structure and terminology, return to decisions made dozens or hundreds of pages earlier, and translate and transcreate works into other languages.

What had once required fluency in other languages, a large team, and money I did not have suddenly became practically achievable.

At first, it felt almost like liberation.

I already had a large audience on social media — tens of thousands of followers. That made it possible to test the result with other people, not only by myself.

I gave translated texts to people who knew the target language well and asked them to read them and tell me how natural they sounded, where they felt machine-made, where there were errors or strange passages, and where the text read like normal literary prose.

In effect, these were my first informal focus groups.

And gradually it became clear that AI was already capable of producing text that a native speaker of another language could judge to be good, natural, and acceptable as literature.

But that was when a question appeared that I had never had before.

A person can tell me that the translation is good.

But do they know what, exactly, was important to preserve?

And did I myself remain in this flow?

Was the book on the other side still my book?

My thoughts.

My understanding of the human being, life, and reality.

My doubts and contradictions.

My way of looking at things.

What I actually wanted to say — rather than a more convenient, smoother, more understandable version of it.

And then I understood that the question was no longer only about translation.

For me, Voiceprint began with that question.

From a Personal Question to Research

The personal story gave us a starting point. But Voiceprint is not a study of one author.

We are exploring a broader question:

what happens to authorial presence when a text passes through another language and a machine?

AI can preserve facts.

It can convey the general meaning correctly.

It can make a sentence sound natural to a speaker of another language.

It can smooth out roughness, remove repetition, explain ambiguity, replace an unusual turn of phrase with a more familiar one.

But sometimes what belongs to the author is precisely what those “improvements” remove.

That is why it is not enough for us to ask:

“Is the text translated well?”

We have to ask a different question:

“What must be preserved for this text to remain the text of its author?”

02

Loss

A Correct Text Can Still Be Wrong

Modern language models are exceptionally good at normalization.

They can turn the strange into the familiar.

The heavy into the light.

The ambiguous into the clear.

The uneven into the smooth.

The unusual into something statistically more expected.

In many tasks, this is an advantage.

In an authorial text, it can become a loss.

Because a human voice is not made only of beautiful sentences.

Sometimes it is located precisely in the place an ordinary editor or a model would want to correct.

Translation and Transcreation

The usual question of translation sounds something like this:

“How do you say this in another language?”

For Voiceprint, that is not enough.

We are interested in another question as well:

“What is happening here — and what must happen in another language for the text to remain itself?”

That is why we do not treat transcreation as permission to rewrite an author freely.

Quite the opposite.

The more freedom a tool receives, the more important it becomes to understand the boundaries of that freedom.

A change does not become authorial simply because it works well.

It becomes authorial only when the right to make that change remains with the author.

What Is Authorial Presence?

Authorial presence cannot be reduced to a single style.

It is not a set of favourite words.

Not the frequency of long sentences.

Not the ability to teach a model to “write like this.”

For us, it is a system of decisions.

It can include:

meaning and intention;
rhythm;
terminology;
recurring images;
metaphors;
structure;
the temperature of dialogue;
the degree of harshness or softness;
deliberate understatement;
ambiguity;
strange formulations;
intentional roughness;
and even the author’s right not to be improved.

And far from all of this can be measured automatically.

Fluency Still Does Not Mean Fidelity

One of the first problems that became visible in this work was this:

fluency ≠ fidelity

A text can read beautifully in the target language while gradually becoming a different text.

This is an especially dangerous kind of loss because it is almost invisible.

An obvious error can be seen.

A missing sentence can be found.

A mistranslated term can be corrected.

It is much harder to notice the moment when a text is still saying roughly the same thing — but is already looking at the world a little differently.

03

Author and AI

The Human Remains the Source of Meaning

Voiceprint is built around a simple boundary:

human-authored / human-directed / AI-assisted

AI can help.

It can suggest options.

Compare.

Find discrepancies.

Check terminology.

Remember decisions.

Point to places where meaning may have been lost.

Help an author work with a language they do not know well enough themselves.

But AI does not get the final say in what a work is.

Complexity can be delegated. Canonical authority cannot.

ATP — A Methodological Lineage

Voiceprint is connected to a separate, mature methodological lineage — ATP, Ashraellen Transcreation Protocol.

ATP began as a practical system for working with literary transcreation and, over time, developed into a separate formalized methodology with its own history, versions, public repository, and DOI.

Its central question is very close to Voiceprint:

How do we detect that the author is beginning to disappear from the transformed text?

That is: how do we detect the moment when the author begins to disappear from the transformed text?

ATP does not promise a perfect translation.

Its purpose is to make decisions visible, preservable, and verifiable.

Memory Matters More Than a Good Prompt

This work leads to another important conclusion: the problem of authorial presence cannot be solved with one very good prompt to a model.

A work hundreds of pages long creates a history of decisions.

Why was this term chosen this way?
Why was the roughness left here?
Why does this character use this particular word?
Why was the ambiguity not removed?
Why was a similar sentence thirty pages earlier handled differently?

So reproducibility, for us, does not mean:

“the model should produce the same text every time.”

It means being able to reconstruct a system of authorized decisions.

The durable object is not the prompt.

The durable object is the system of decisions around the work.

04

Method

MONOLITH Is Not a Laboratory Example

The MONOLITH trilogy became an important environment for this work.

It exists independently of Voiceprint and has its own authorial canon.

We did not create it as a demonstration corpus for the research.

Working with real long-form works made certain problems especially visible — problems that later became part of the research work around Voiceprint and ATP.

A long text quickly reveals what is almost impossible to see in a single paragraph.

A decision made today can change the meaning of a scene two hundred pages later.

A name. A repetition. A metaphor. A character’s intonation. A philosophical term.

In a long work, translation has to remember.

What Should AI Leave Untranslated?

One of Voiceprint’s questions can be put even more radically:

What must an AI not translate?

What should the machine leave alone?
What must not be normalized automatically?
When should the strange remain strange?
When must ambiguity not be turned into explanation?
When is an uncomfortable sentence not an error, but part of the author?
When does a more beautiful version become less faithful?

For now, we do not assume that there is one universal list of answers.

But the question itself is fundamental for us.

Because a good system should understand not only what it can change, but also what it has no right to change on its own.

Research Must Be Allowed to Fail

Voiceprint is not built around the need to prove a conclusion chosen in advance.

Perhaps some elements of authorial presence can be preserved reliably.

Others only partially.

Some may prove too subjective to formalize.

Some methods may not work.

ATP may need to change.

Individual hypotheses may be rejected.

For us, that is a normal result of research.

If a system does not allow for the possibility of its own failure, it is no longer research.

Presence Cannot Be Reduced to a Single Number

There is a temptation to imagine that one day we might say:

Author Presence: 93%

But today we have no basis for claiming that human or authorial presence can honestly be measured by one universal metric.

Individual features can be measured.

Terminological stability.

Structural preservation.

Semantic divergence.

Consistency of decisions.

Reader evaluations.

But turning all of this into one elegant number would be premature.

So for now, Voiceprint is looking more for a system of loss indicators than for a magical index of presence.

05

Living Voice

From Text to a Living Voice

Books were the first environment in which this question became obvious to us.

But the question itself is broader than literature.

A person speaks in one language.

A machine helps carry that speech into another.

Another person hears the message.

And the same question appears again:

who arrived on the other side?

Only the information?

Or the person who was trying to convey it?

This is where Voiceprint Live emerges as a direction of research.

We are exploring the possibility of carrying live speech across languages in a way that attempts to preserve — as far as this is possible and verifiable — intention, context, register, and the individual manner of expression.

This is still a developing direction, not a promise of finished technology.

Not a Digital Twin

Preserving human presence does not require creating a digital copy of a person.

In Voiceprint’s current architecture, the human remains the subject.

The system may help their speech pass through a technical and linguistic barrier.

But it does not gain an independent right to carry on that person’s presence in their stead.

For us, this is an important boundary.

Not Everything Should Become Universal

Large language models are very good at reproducing familiar and probable ways of expression.

But human culture often emerges precisely where a person speaks not in the most probable way.

In strangeness.

In imperfection.

In local expression.

In an error that has ceased to be an error.

In an intonation that cannot be optimized without loss.

That is why we are interested not only in AI’s ability to cross language boundaries.

We are interested in the price of crossing them.

Where Am I in This Flow?

We began with a book.

But the question turned out to be much larger than a book.

If human thought passes through an increasing number of machine transformations, it is not enough to ask how good the final result is.

We need to be able to ask:

What was preserved?
What changed?
Who decided that?
Can the history of that decision be reconstructed?

And where is the boundary between helping a person and quietly replacing them with a more convenient version of themselves?

Voiceprint does not yet claim to know the final answer.

That is our research.

What Must Remain

AI can give a person enormous new freedom.

It can help a book cross a language.

Thoughts — cross a cultural boundary.

A conversation — pass through misunderstanding.

A person — reach places they might once have been unable to reach at all.

We do not want to stop this flow.

We want to understand what must not be lost inside it.

A machine can help a person go farther.
But on the other side, the one who began speaking must still be there.