For our first year, MulingStream translated with traditional machine translation, the kind that works through a sentence piece by piece. On 19 March 2026 we moved to large language models, which understand the meaning of a sentence in its context before they say it again in another language. Here is why, and what it changes for the person listening.
Every word was translated, and the sentence still made no sense.
That was the problem we kept running into in our first year. MulingStream translated with traditional machine translation, the kind built on classic natural language processing. It was fast and it was steady, and for plain sentences it did well. But a sermon is not made of plain sentences.
A pastor says "he took up the cross". A tool working piece by piece does not know whether that is a hill, a punch in boxing, or Calvary. It picks one, and it picks the same wrong one in Arabic, in Polish and in Tamil, at the same moment, and nobody in the room can tell you it went wrong.
The same thing happens a dozen times in an ordinary talk. Words with many meanings, such as spirit, grace or charge, where the right one depends entirely on what is being talked about. Sentences that restart halfway: "so what Paul is saying, well, what he is really saying is". Sayings that mean nothing taken literally, like the Danish "there is no cow on the ice", which simply means there is nothing to worry about. Translated word for word, the guest gets a cow.
A listener can live with one odd sentence. After twenty minutes of stiff word order and phrases nobody says out loud, people stop listening. That is the difference between following a service and enduring one.
On 19 March 2026 we moved MulingStream's translation to large language models, often called LLMs. The difference is in what happens first. Traditional machine translation starts from the words. A large language model starts from the meaning: it understands what the sentence is doing in its context, what the talk is about and what has already been said, and then it says it again the way a person speaking that language would say it.
We call it Smart Translation, and it is how MulingStream translates, for every room, from the first sentence. There is nothing to switch on.
It translates what you meant, not the words you used.
The guest is no longer decoding a chain of swapped words. They are simply listening, the same as everybody else in the room.
Church words keep their church meaning. Many languages have two words for the same thing: an everyday or clinical one, and the one the Bible in that language has used for generations. A meaning-first translation reaches for the one that sounds like scripture, so the sermon sounds like a sermon.
Sayings arrive as what they mean. "There is no cow on the ice" comes out in English as "there is no need to worry", and in Persian and Ukrainian as the same reassurance, with no cow in any of them. You can see real examples from Danish, English, Spanish and Persian on the Smart Translation page.
It reads like somebody wrote it. Whole sentences, in the word order that language actually uses, so a guest can relax into the service instead of working through it.
Not in a way the room notices. What changes is the rhythm: instead of words arriving one at a time, the guest gets a whole sentence once its meaning is clear. That is easier to follow, not harder, and some languages need it. In German, for example, the word that decides a sentence often comes at the very end.
Understanding meaning is exactly right for most words, and exactly wrong for a name. Grace, Joy and Hope are ordinary words, and they are also somebody's name in your congregation. That is what Keywords are for: the names, places and community words that must come through untouched, in every language at once.
And when a reference needs more than a translation, the verse a pastor names by its number, or a saying everybody local knows, Smart Hints add one short line of explanation beside it. We released them the same day, and you can read why we built Smart Hints.
Smart Translation is not an upgrade and not a box you tick before the service. It works in every language the room is running, it needs no preparation and no training, and nobody has to watch it during the service. It is included for everyone, whichever way you pay. Choose from 200+ languages on the languages page.
After a season of services, a church in Germany wrote to us about exactly this: the short wait for the whole sentence.
Some commented on the delay between original sentence and translated version. But when I explained the ‘context based’ concept of the app, [they] immediately agreed. In German preaching and talks (…) we love to create long and complex sentences, and we often have crucial keywords pretty much at the end of a paragraph. So the ‘context based’ concept is very useful and helps to improve the quality a lot.
The best way to judge a translation is to hear it. Open the listener demo room, pick your language, and listen to a real talk.
We show you the whole thing live, in about fifteen minutes. Nothing is decided on the call and nothing follows it: you hear it, and then you are the judge.