Prof Dr Peter Gilles from the Institute of Luxembourgish Linguistics at the University of Luxembourg has for years been working on making parliament more accessible to Luxembourg's diverse population.
While French remains the administrative language and is used for all official and legal texts in Luxembourg, much of the debate in the Chamber of Deputies is held in Luxembourgish.
"Considering that this represents only 50% of the population who can understand", Gilles explains in conversation with RTL Today, "there is clearly a need to offer these public decisions, organisation, and institutions to the entire population."
It was only around 10 years ago that the possibilities changed entirely with the arrival of deep learning algorithms. And with that in 2022, the professor, whose academic background is in computational linguistics, finally found an avenue to realise his goals.
"I found the possibility to collaborate with the Chamber of Deputies to develop [deep learning for Luxembourgish] further", he said. "Especially for a small language. All these small languages are not left behind, but it is more difficult to develop the systems for them."
Speaking to RTL Today, doctoral researcher Nils Rehlinger elaborated: "Google Translate can translate from Luxembourgish, into Luxembourgish. In terms of quality, if you compare it to bigger languages like German, French, English, the quality is lagging behind. So, we're trying to improve this."
An ambitious and multi-step initiative by the University of Luxembourg, funded and developed in collaboration with the Chamber, was eventually launched in 2022.
"We first developed our system, LuxASR, for transcribing Luxembourg speech", said Gilles. This technology is already being used in the Chamber: for everyday work internally and to transcribe meetings and speeches.
The way a Large Language Model (LLM) is trained in Luxembourgish is the same as with any other language: on a vast amount of text for natural language processing tasks.
LuxASR is trained to a large part on speeches from Parliament, as well as data from RTL, local communes, and radio stations such as Radio 100.7. But all that data is still insufficient to train a good model.
"We don't create the models from scratch because that would require huge amounts of resources", said Rehlinger. "[Instead] we let bigger and better models translate the data, and then we give it to our model so that our smaller model can still improve a bit."
After years of training and almost 50 models developed and improved over the years, LuxASR is ready for production.
President of the Chamber of Deputies Claude Wiseler provided us with further insights on the adoption of the new technology so far.
"We use it for our internal purposes to facilitate the work we have to do and just to schematise – to make a transcription of what is said in Parliament", he noted, pointing to the constitutional obligation to document every word said in public sessions.
Publicly available parliamentary discussion and debate is not something new, it has been done manually since the 19th century.
During commission sessions, the AI transcription technology is also used for taking notes or summarising discussions. In this case, the AI-generated transcriptions can be checked for grammar and accuracy after they are produced.
With the technology already ready, the next step is near completion: to translate Luxembourgish dialogue into Luxembourgish text in real time.
"Putting subtitles online will help us with two things: first, of course, with people who have hearing difficulties, and because a lot of people watch TV without [audio], on the bus or train, or on their phone. And they [could] read subtitles without having to listen", said Wiseler.
Another benefit of Luxembourgish audio-to-text subtitles is that it can help those who know some Luxembourgish, but have not yet reached full fluency. "Reading is easier [than listening], so people with a low knowledge of Luxembourgish are better able to follow if they have the help of subtitles", the Chamber president added.
The ultimate goal is more ambitious still: LuxMT, a model being developed to automatically transcribe, document, and translate live-streamed Luxembourgish parliamentary debates by providing multilingual subtitles in real time – in French, German, Portuguese, and English.
As good as a model can be, when faced with something as consequential as parliamentary debate, there are, without question, risks. In the case of LuxMT, the question is where the line for risk tolerance is.
"Whatever you use AI for, in our case transcription or translation, you need to do a risk assessment. So, how risky is it to have mistakes? Will there be big consequences for mistakes, or is it okay if there are small mistakes?", Rehlinger said.
"If you're in a situation where there's a high risk, for example, medical or the law, where there will be very real repercussions, then, of course, there needs to be a human in the loop."
For LuxMT, it is the speed and quantity of material being worked on that makes it suitable for AI use – but the team acknowledges its limitations.
"No model, not even for English, or the best one, is 100% reliable", Gilles stressed. "Parliament, of course, is quite a sensitive issue and you could create severe misunderstandings if it is real time, because you cannot correct it – it is on the TV, it is in the world."
More challenges present themselves through specific words and phrases a machine cannot know.
"And that will continue to be the case, especially idiomatic expressions, names, brand names, specific institutions, names for institutions, which are very specific for a country, for example", Gilles added. "This, I think, can only be solved by humans.”
Rehlinger shared that the next steps are working with professional translators: "To really focus on language that's very specific to the Luxembourgish parliament, so all the terminologies, the special terms related to the law and so forth – to really specialise the model for this use case."
One solution could be a delayed response, where a human acts as a middle person to control the output of the translation and make corrections. Another is what most of modern society has already become accustomed to: understanding that real-time captions serve a purpose, rather than strive for perfection.
Growing up in Luxembourg and appreciating diverse linguistic backgrounds among his peers from a young age, Rehlinger considers himself an advocate for multilinguality and sees this project as a real opportunity to improve accessibility in Luxembourg.
"The topic of multilingualism in Luxembourg to me always is quite often split between advocating for Luxembourgish and advocating for multilinguality", he said. "I find it a bit odd that it seems to be difficult to do both. Multilingualism is part of what makes Luxembourg so special."
Gilles added that the country’s trilinguality and a strong international presence reassert the need for greater accessibility.
"It is only through flexibility that we as a Luxembourgish nation can define an optimal or workable way of living together and using multiple languages", he said.
For Wiseler, the reality of real-time translations in Parliament is an exciting and encouraging one.
He explained: "If you live in a country, you should have access to what's happening. Even if [residents] can't vote or don't want to vote, they will be concerned. And if you are concerned, you have the right to understand."
What is also important, he stressed, is integration.
"We want people to integrate in Luxembourg, to feel at home, and at some point of their life to take Luxembourgish nationality and to be interested in what's happening, being interested in politics", Wiseler said. "I think it's important to at least make the offer to them to understand."
While the possibilities continue to grow, the first step for using this technology is clear for Parliament.
"What we are trying to do in Parliament [is] having an AI instrument to make subtitles in Luxembourgish for the parliamentary debates", Wiseler said. "We are almost ready to do that. The next step will be to have the automatic translation in French and English at least, so that you can follow on YouTube: Parliament with subtitles."
Alongside the ongoing work of the University, more is being done still.
A bot that allows Q&A and conversation in Luxembourgish is being developed. Currently, a smartphone application that transcribes Luxembourgish voice messages into Luxembourgish text is available for Android and iOS. A web version is also available for use.
Later in 2026 or at the start of 2027, the team also plans to launch a survey for rating the AI translations, which will provide more information on what needs further improvement.
