Why AI Interpretation Still Gets Names Wrong After Preparation

Event organizers who use Nubart TRANSLATE, our AI simultaneous interpretation system for live events, send us their programmes, slides and speaker lists. Our team uses these materials to prepare the AI for their event. Occasionally, one of our customers will write to us afterwards asking, "The company name was in my slides, so why did the AI get it wrong?" It's a fair question. More often than not, the mistake occurred before any translation began: the system heard something other than what the speaker said. As a conference interpreter, I know that problem well. A name that you read in the program can still sound like something else when a speaker says it quickly. This article explains the three steps in the AI interpretation process where names can go wrong, and how preparation of your documents can help.

---

Three steps, three places where a name can go wrong

AI interpretation involves three systems working sequentially. First, speech recognition listens to the speaker and transcribes what is said. A translation engine then translates the text, one short sentence at a time. Finally, a synthetic voice reads the translation aloud. Nubart TRANSLATE provides every translation in both audio and text formats.

1. 1. Speech recognition writes down what the speaker said Typical error: misheard
2. 2. Translation translates the text, sentence by sentence Typical error: wrong meaning
3. 3. Voice reads the translation aloud Typical error: mispronounced

A name can go wrong at each of these steps, and a different fix is needed for each one. A customer who tells us that a term is incorrect always thinks "translation." But after hundreds of events, we know better and think "transcription!" Our first question is always the same: What did the speech recognition program write down? If the speech recognition software wrote down the wrong word, the translation engine never saw the right one.

When the system hears something else

Take the company name Accenture, spoken by a German speaker in a German presentation. If the name appears in the documents, our team prepares the speech recognition for it. Essentially, we tell the system that if it hears something that sounds like "aksentür," "axentür," or "aksentschur," the speaker means Accenture, and it should write Accenture. Then, the translation engine receives the correct name from the transcription, understands it and keeps it.

Now a speaker pronounces it in a way nobody expected, closer to "Achsentür". That happens to be a German word: an axle door (whatever that means!). If this sound was not among the prepared variants, the speech recognition may write "Achsentür", and the translation engine translates it faithfully. English listeners hear about an axle door.

The translation is a correct translation of the wrong word.

The same happens with the names of people. A French name like Thibaut sounds quite different when an English, Spanish or German speaker says it. We prepare the plausible variants. But we cannot prepare every pronunciation that every speaker might use, with every accent and at every speed. Based on our experience with Nubart TRANSLATE, the speech recognition not "understanding" the speaker is the reason for three quarters of all mistakes.

Research points the same way. A study on person names in speech translation found that these systems handle names poorly, and that in systems which first write down the speech and then translate the text, the errors came largely from the speech recognition (Gaido et al., 2022). A more recent evaluation from 2024 tested speech translation systems on names in recorded talks. Giving the systems a list of the names that would come up improved the results, but even the best system still missed about one in eight organization names (Awiszus et al., 2024). Those systems could process the whole recording at once, while live interpretation has to decide in real time. Both findings match what we see when we check our own transcripts.

When a rare term sounds like a common one

A second case is harder. Suppose a company has a Strategic Advisory Board that everyone inside calls the SAB. Spoken aloud, SAB and SAP differ by a single sound. The software company SAP appears far more often in the language these systems learned from than any company's internal board. When the sound is not perfectly clear, speech recognition tends to choose the more familiar word.

Preparation pushes the system toward SAB, and it often helps. But we cannot tell it that everything that sounds like SAP is really SAB, because a speaker at the same event may well talk about SAP. Short abbreviations are especially exposed to this, and so are names that sound like ordinary words.

When the word was heard correctly

Sometimes the speech recognition writes the right word and the mistake happens later.

The translation engine works one short sentence at a time and does not know what the talk is about. A word with two meanings can come out with the wrong one. The German word "Schalter" is a switch in an engineering talk and a counter in a bank. For the translation engine, our team writes a short context text about the event in the speaker's language. The engine reads it with every sentence, much as an interpreter keeps the briefing notes in mind. Where a term must always be translated one way, or never translated at all, our team sets that by hand for each listener language. If the context is technical, "Schalter" will most likely be translated as "switch" and not as "counter".

The last step is the voice. A synthetic voice reads a foreign name by the rules of its own language, so an English voice may turn the German surname Grünwald into something the speaker would hardly recognize. Here too, the pronunciation can be set by hand.

Names also need different treatment depending on what they are. A company name like Accenture stays the same in every language. A city may have its own name in another language: Köln becomes Cologne in English, and that is correct. A person's name stays the same in writing but may need a different pronunciation for each voice. An abbreviation is sometimes spoken as a word and sometimes letter by letter. This is why a simple glossary rarely solves the problem: each name needs the treatment that fits it.

For the most relevant names, our team looks up the pronunciation in the International Phonetic Alphabet (IPA) and applies it, based on the speakers' languages, which the customer tells us when placing the order. About 60% of our preparation process is automated and works on the files the customer uploads to the password-protected customer dashboard. Decisions like these, however, are made manually by our linguistically skilled team, not by the AI.

Words that were never prepared, and why more is not better

Preparation works on what the material shows. If a name does not appear in the documents, our team cannot know that it needs special treatment. The same applies to material that arrives too late. Some parts of an event cannot be prepared by anyone, because events are live and unpredictable by nature. In a Q&A session, nobody knows what the audience will ask or which names the answers will contain. And the best speeches tend to be the spontaneous, unprepared ones. A human interpreter faces the same problem. What this means for your guests, and for the agreement with your own client, I explain in Can AI Interpretation Make Serious Mistakes?

The natural reaction of many customers is to send everything: every document, every term list, hundreds of pages. That usually does not help. Our team prepares the speech recognition only for the words that are really at risk: names, abbreviations, coined product names, unusual spellings. Ordinary words stay out on purpose, because a long list can start to affect how the system hears the ordinary words around the listed ones. The context text for the translation engine has a length limit for a similar reason. It is sent with every sentence, and an overloaded context can make the translation worse. Choosing what to leave out is as much a part of preparation as choosing what to put in. I describe the same effect for term lists in What to Expect from AI vs. Human Interpretation.

Preparing the AI is different from training it

Some customers assume that the AI learns their terminology and gets better with every event. It works differently. In the context of AI translation, the word "training" implies a model that would have to be fed with huge amounts of data. This is feasible, but out of our scope, and definitely an expensive overkill for a single event. Instead, our team "prepares" existing models specifically for this event, as you would prepare a human interpreter who already has "a model" in their brain and only needs to know what the event will be about. The AI model underneath stays exactly as it was, and nothing from your event is kept in the model afterwards.

This has two consequences. Your material serves your event only: event content is not used to train AI models, and our partners are contractually prohibited from doing so. And the system does not improve by listening. Reading a presentation aloud several times before the event makes no difference to the result. Each event is prepared from its own material.

What helps, and what to do when a term went wrong

Send us what you would give a human interpreter: the programme, slides, speaker biographies, product sheets. We know what to do and we will take care of it. Make sure the names of the speakers, the products and the venue appear somewhere in them. You do not need to build a glossary. If some names or terms must be translated or pronounced in one particular way, add a short list of just those. Send everything as early as you can, because material that arrives in the last hours before the event may not make it into the preparation.

If a term went wrong at your event, tell us which word you expected, roughly when it was said and in which room. Unless you chose immediate deletion of your event data, we can check the transcript and see at which of the three steps it went wrong. Then we know whether the fix belongs to the speech recognition, the translation or the voice, and we can include it when we prepare your next event.

For even more peace of mind, you can request a test with a prepared AI. For this paid option, our team prepares the system for your event, and you can try it for 30 minutes. If some words still do not come out right despite the preparation, let us know and our team will work on them again. One round of revision is included.