AI That Speaks Lithuanian — How We Built It

When I first heard our AI correctly decline a Lithuanian surname in the vocative case — addressing a caller as "Sveiki, Jonai" instead of the grammatically wrong "Sveiki, Jonas" — I knew we were onto something. That single inflection, a vowel change at the end of a name that most global AI platforms get wrong every single time, represented months of work. And it was just the beginning.

This is the story of why we built AINORA, how we approached the enormous challenge of making AI speak Lithuanian naturally, and what we learned along the way. It is a story about language, technology, and the stubborn conviction that three million Lithuanian speakers deserve AI that actually understands them.

Why Lithuanian Is One of the Hardest Languages for AI

If you are reading this in English, you are used to a language where words barely change form. "John" is always "John" — whether you are talking to John, about John, or giving something to John. Lithuanian does not work that way. Not even close.

Lithuanian is one of the most archaic living Indo-European languages. Linguists study it to understand how ancient languages sounded thousands of years ago. For AI engineers, this ancient heritage translates into a very specific set of challenges:

Seven grammatical cases

Every Lithuanian noun, adjective, and pronoun changes form depending on its role in the sentence. The name "Jonas" alone has seven forms: Jonas (nominative), Jono (genitive), Jonui (dative), Joną (accusative), Jonu (instrumental), Jonui... wait, that was dative. Let me be precise: Jonas, Jono, Jonui, Joną, Jonu, Jonuje (locative), Jonai (vocative). Each form signals a different grammatical function. Get the case wrong and you sound like a foreigner who learned Lithuanian from a textbook — or worse, like a machine.

For a voice AI that needs to greet callers by name, this is critical. When a customer named Jonas calls, the AI must say "Sveiki, Jonai" (vocative), not "Sveiki, Jonas" (nominative). When discussing Jonas with a colleague, it would be "Jono vizitas" (genitive). This is not optional politeness — it is the basic grammar that every Lithuanian speaker internalizes by age five.

Complex declension patterns

It is not enough to know that names change. You need to know how they change, and Lithuanian has multiple declension patterns. "Jonas" becomes "Jonai" in vocative, but "Petras" becomes "Petrai." "Kazys" becomes "Kazy." "Giedrius" becomes "Giedriau." Female names follow entirely different patterns: "Ona" stays "Ona" in vocative, but "Rasa" becomes "Rasa" while "Jurgita" becomes "Jurgita." Then there are surnames: "Butkus" becomes "Butkau" in vocative, "Petrauskienė" stays "Petrauskiene."

Now multiply this by every noun in the language — not just names but places, services, time expressions, and everything else that appears in a business phone conversation. The combinatorial complexity is staggering.

Formal and informal registers

Lithuanian has a strong distinction between formal ("jūs") and informal ("tu") address. A business AI must always use the formal register with customers — but this formal register affects verb conjugations, pronouns, and even some noun forms throughout every sentence. A single slip into informal register can sound rude or unprofessional.

Pronunciation and intonation

Lithuanian has pitch accent — the same syllable can be pronounced with a rising or falling tone, changing the meaning. City names like "Vilnius," "Kaunas," and "Šiauliai" have specific stress patterns. Lithuanian surnames have stress on different syllables depending on the declension class. Mispronounce a customer's name and you immediately lose credibility.

Small global dataset

Here is the fundamental problem: modern neural speech models learn from data. English has billions of hours of training data. Lithuanian has a fraction of that. With approximately three million native speakers and limited digital content compared to major world languages, Lithuanian is what AI researchers call a "low-resource language." Global platforms train primarily on English and the top ten world languages. Lithuanian is an afterthought — if it is supported at all.

The Problem We Set Out to Solve

In 2025, when I started exploring voice AI for Lithuanian businesses, I tested every major platform I could find. The results were, to put it diplomatically, not ready for production.

I called a demo of a well-known global voice AI platform and tried to interact with it in Lithuanian. It understood maybe 60% of what I said. When it tried to speak Lithuanian back, it sounded like someone reading a Lithuanian text with an English accent — stress on the wrong syllables, vowels flattened, cases jumbled. It called me "Justas" when it should have said "Justai." It pronounced "Vilnius" as something closer to "Vil-nee-us" instead of the correct "Vil-nius" with the proper soft "l."

For a hotel trying to impress guests, or a dental clinic trying to build trust, this was worse than useless. A voice that sounds foreign and makes grammar mistakes does not build confidence. It actively damages the brand.

I realized that no one was going to solve this problem for us. Global AI companies optimize for English and Mandarin and Spanish — languages with hundreds of millions of speakers and massive commercial markets. Lithuanian, with its three million speakers and GDP-sized market, was never going to be a priority. If Lithuanian businesses were going to have voice AI that actually works, someone who understood both the language and the technology needed to build it.

That someone became us.

What Goes Wrong When You Just “Add Lithuanian”

Before I explain how we solved it, let me show you what happens when a global AI platform simply "adds Lithuanian" as one of its supported languages. These are real examples — the kinds of errors we encountered in our testing:

Every single one of these errors is noticeable to a native Lithuanian speaker. And they compound. After hearing two or three wrong cases in a row, a caller stops trusting the AI. They ask to speak to a human. The whole point of the system — handling calls efficiently and professionally — is undermined.

How We Built Lithuanian AI That Actually Works

Building truly Lithuanian-speaking AI was not a single breakthrough. It was a process — methodical, iterative, and often frustrating. Here is how we approached it:

The name problem — our biggest challenge

Lithuanian names were our single biggest technical challenge. Consider what the AI needs to do when someone calls:

1. Recognize the caller (by phone number or by asking their name)
2. Determine the correct vocative form of their first name
3. Use the correct genitive form when discussing their records ("Jono vizitas," not "Jonas vizitas")
4. Handle surnames correctly in different contexts
5. Pronounce everything with the correct stress pattern

Each step requires different linguistic knowledge. And the edge cases are endless: What about the name "Tomas" — vocative "Tomai"? What about "Lukas" — vocative "Lukai"? What about "Dovydas" — vocative "Dovydai"? These all follow slightly different patterns despite looking similar.

Then there are foreign names that have been adopted into Lithuanian: "Robertas" (vocative "Robertai"), "Edvardas" (vocative "Edvardai"), "Kristina" (vocative "Kristina"). And names from other cultures entirely — what do you do when a caller named "Ahmed" or "Wei" calls a Lithuanian business? The AI needs to handle these gracefully too, recognizing that not all names follow Lithuanian declension patterns.

We built a name processing system that handles thousands of Lithuanian names with their correct forms, recognizes foreign names and treats them appropriately, and falls back gracefully when encountering a name it has not seen before. It took months. It was worth it.

Beyond names — getting the whole conversation right

Names were the most visible challenge, but correct Lithuanian extends far beyond names. Consider a simple appointment confirmation:

“Jūsų vizitas pas gydytoją Petrauskienę yra trečiadienį, vasario dvidešimt aštuntą dieną, keturiolika valandą trisdešimt minučių.”

In that single sentence, the AI needs to: use the accusative case for the doctor's name after "pas" (Petrauskienę, not Petrauskienė), use the accusative for the day of the week (trečiadienį, not trečiadienis), correctly decline the ordinal number for the date (dvidešimt aštuntą, not dvidešimt aštuoni), and use the accusative for the time (keturiolika valandą, not keturiolika valanda). Four different case decisions in one sentence, all of which must be correct or the whole thing sounds wrong.

The Results: Global Platform vs. AINORA Lithuanian

After months of development, testing, and iteration, the difference between generic Lithuanian support and what we built is dramatic:

But numbers do not tell the whole story. The real test is this: when a Lithuanian person calls a business using AINORA, they do not immediately realize they are talking to AI. Not because we are trying to deceive them — the AI identifies itself at the start of every call, as required by the EU AI Act — but because the Lithuanian sounds natural. Correct cases. Proper intonation. Natural phrasing. The kind of Lithuanian you would expect from a competent, professional receptionist.

That is the standard we set for ourselves, and that is what we deliver.

Where We Are Now

Today, AINORA handles calls for dental clinics, hotels, beauty salons, veterinary clinics, auto service centers, and other service businesses across Lithuania. Every one of those calls happens in natural Lithuanian — with correct grammar, proper name handling, and professional intonation.

But we are not done. Language is a living thing, and Lithuanian is evolving. New names appear, new business terminology emerges, regional expressions shift. Our evaluation pipeline catches regressions, our linguist consultants review edge cases, and our system gets better with every call it handles.

We are also expanding what our Lithuanian AI can do. Beyond answering calls and booking appointments, we are building capabilities for outbound call automation, customer memory and personalization, and deep CRM integration — all in native Lithuanian.

The bigger picture is this: Lithuanian businesses should not have to settle for AI that sounds foreign or makes grammar mistakes. Three million people speak this beautiful, ancient language, and they deserve technology that respects it. That is what we are building, and we are only getting started.

What this means for Lithuanian businesses

If you have been hesitant about voice AI for your Lithuanian business because you tried a global platform and the Lithuanian was not good enough — we understand. We felt the same frustration. That frustration is exactly why AINORA exists.

Your customers call expecting to be greeted in proper Lithuanian. Your brand depends on professional communication. You cannot afford an AI that calls your customer "Jonas" when it should say "Jonai," or that mispronounces your city's name, or that stumbles through case declensions like a first-year language student. See how our technology works to understand the engineering behind natural Lithuanian voice AI.

We built AINORA specifically so you do not have to worry about any of that. The Lithuanian just works — because we made sure it does.

Frequently Asked Questions