Giving Luxembourgish a Voice in the Digital World

Publié le 20 novembre 2026

What if the future of Luxembourgish depended not only on the people who speak it, but also on the data we make available to technology? Interview with Daniel Wagener, linguist at ZLS.

From translation tools to AI-powered applications, the Zenter fir d’Lëtzebuerger Sprooch (ZLS) is opening up new possibilities to ensure that Luxembourgish has a place in the digital world.

The ZLS recently published new datasets on the Open Data portal: could you introduce them to us?

As one might expect, the various datasets that the Zenter fir d’Lëtzebuerger Sprooch (ZLS) has published over the course of the year on the Open Data portal all pertain to the Luxembourgish language. However, they span various formats and serve different purposes.

The first dataset, produced as part of the Tech-in-GOV initiative 2025 and released in mid-2026, is a multilingual translation corpus consisting of 150,000 Luxembourgish source words translated professionally into English, French, and German. The source text is composed of various resources that are either in the public domain or that the ZLS obtained explicit permission to use. It includes, among other sources, example sentences featured in the Lëtzebuerger Online Dictionnaire (LOD), news articles, and transcriptions of parliamentary debates. The source text was checked for spelling and grammatical errors, and the translations were carried out by professional translators.

We also published a substantial portion of the training data behind Schreifmaschinn, our speech-to-text engine that is part of our platform Sproochmaschinn, in two separate datasets. Most of the data is released under a CC0 licence, with a smaller portion available under CC BY. Together, both subsets make up a total of 60 hours of transcribed Luxembourgish audio from news broadcasts, parliamentary debates, podcasts and other sources. We opted to publish two distinct versions of the transcripts: one that reproduces the audio verbatim, including hesitation markers, repetitions and other speech disfluencies, and another in which these features have been normalised or filtered out. The verbatim version offers a representative insight into spoken Luxembourgish, while the normalised version is closer to written Luxembourgish.

Finally, we published the complete word list behind Spellchecker.lu, our online Luxembourgish spellchecker. It contains an even more extensive collection of orthographically correct Luxembourgish words than the vocabulary covered by the LOD, making it a valuable resource for developing digital applications for Luxembourgish.

The ZLS is a true champion of Open Data: what is the primary motivation behind this openness policy?

This policy is rooted in both practical considerations and a broader principle. As a public institution, the ZLS is committed to making its linguistic resources available whenever possible and in accordance with applicable legal requirements. At the same time, the policy reflects our conviction that the ZLS’s work is funded by public money and that its results should therefore be accessible to the public wherever feasible.

Our commitment to open data is also motivated by the fact that we want Luxembourgish to keep pace with developments in language technology, an aspect that has become increasingly important in recent years. Whether we are talking about relatively simple applications such as spellcheckers or more sophisticated AI-driven tools such as machine translation engines, all of them rely on data to function properly. For a language as small as Luxembourgish, the market alone is unlikely to produce data in sufficient quantity, let alone sufficient quality. At the same time, the ZLS is a small team and cannot develop every application itself. What we can do is provide the high-quality data others need to build such tools.

We therefore make our data available to everyone on equal terms, whether they are students, start-ups, public administrations or multinational companies. We want anyone with a good idea to have the opportunity to put it into practice, without access to linguistic resources becoming a barrier.

Beyond the datasets currently available, what policy does the ZLS intend to pursue in this area in the short and medium term?

Our guiding principle at the ZLS is simple: whenever we produce linguistic data, we aim to make it publicly available to the greatest extent possible. Our approach is based on the idea of being “as open as possible”. We therefore favour the CC0 licence, which allows data to be reused without restrictions. If a partner providing us with linguistic data sets conditions that are incompatible with CC0, such as attribution requirements or restrictions on commercial use, we respect those conditions and publish the material concerned as a separate dataset. This ensures that each dataset carries a single, clearly defined licence.

Making data available, however, is only the first step. For linguistic resources to be genuinely useful, people need to be able to find them, understand them and work with them. We therefore promote our datasets through channels such as the ZLS’s newsletter and LinkedIn, and we provide descriptions in both Luxembourgish and English, together with detailed metadata and files in a range of formats. We want to continue improving the visibility and accessibility of our resources.

Looking ahead, our priorities for the short and medium term are twofold. First, we want to expand our existing datasets, for instance by enlarging our multilingual translation and speech corpora. We are also exploring the possibility of producing Luxembourgish instruction-tuning data to help improve the performance of large language models (LLMs). Second, we want to reach out more directly to developers working on operating systems, software and AI applications to make them aware of the linguistic resources available to them. We also plan to improve the documentation of some of our older datasets, making them easier to reuse. Ultimately, we want to ensure that the linguistic resources produced by the ZLS reach the people who can put them to practical use.

The data currently available has enabled some interesting reuses. What examples have stood out for you?

What is particularly encouraging is the variety of applications being built using ZLS data. Here are just a few examples of the many fantastic reuses that piqued our interest:

  • Mil and Mia, two Luxembourgish voices for screen readers, released by the government’s Information and Press Service (SIP) and based on audio and text resources featured in the ZLS’s Lëtzebuerger Online Dictionnaire (LOD). They allow visually impaired people and those with reading difficulties to have digital content read aloud in Luxembourgish. In our opinion, this is one of the clearest examples of the real-world impact of open data, and one that helps make Luxembourgish content more accessible.
  • LOD to Anki, a project that turns dictionary entries from the LOD into flashcards for Anki, a popular app for learning vocabulary through spaced repetition. Users can build decks tailored to their own level and interests, complete with pronunciation and example sentences, in English, French, German or Portuguese. We particularly appreciate the educational value of this reuse.
  • LOD-MCP, which connects the LOD to AI assistants, allowing them to look up words and verify their use rather than simply guessing. Today, AI assistants still struggle to produce correct Luxembourgish. This project is a step toward addressing that.

Seeing how our data is being used is highly motivating and helps us understand where there is an actual need for high-quality resources in Luxembourgish.

What would you expect today in terms of developments based on your datasets? What proposals do you hope to see come to fruition one day?

One of our key goals is to see Luxembourgish integrated into the digital tools and environments people use every day, as well as those that play an important role in accessibility. Developments we would love to see include more accurate automatic subtitles, Luxembourgish voice input and output on smartphones, as well as spellchecking and predictive text built into keyboards and office software. Speakers of larger languages tend to take these tools or features for granted, but when it comes to the Luxembourgish language, many of them have yet to be developed.

Another area where we would like to see progress is machine translation, particularly regarding the ability to translate both to and from Luxembourgish. The 150,000-word translation corpus we published earlier this year is intended to help support precisely these developments. Similarly, we would like to see AI systems, particularly LLMs, that are able to generate correct and stylistically sound Luxembourgish. Unfortunately, that is not the case today.

Beyond the pure question of Open Data, what is the ZLS currently working on? What projects are in the pipeline, ready to be launched by Christmas or 2027?

A few projects come to mind.

This year marked the release of the eighth instalment in our book series Lëtzebuerger Wuertschatz. Entitled Vun Dräibiz bis Vullemätti – 117 Frechheeten, déi sëtzen, the book is a collection of 117 (mostly) harmless Luxembourgish insults. It is already available in bookshops, and the Christmas season might be a good opportunity to discover it or give it as a gift.

We also launched a hybrid version of our travelling exhibition, D’Lëtzebuerger Sprooch(en). The original exhibition has been touring Luxembourg and its neighbouring countries since 2023. The new hybrid version is available in five languages and can be presented anywhere in the world, both online via its own website, and in print. Instead of having heavy modules shipped from Luxembourg, exhibitors can simply print the necessary materials locally from print-ready files supplied by the ZLS, at a modest cost. This makes it much easier for interested parties to bring the exhibition to their own part of the world.

We are also working on mobile versions of two existing digital tools: the Orthotrainer and the Spellchecker. The Orthotrainer helps users learn and practice Luxembourgish spelling through interactive exercises that gradually build their skills. The Spellchecker comes in handy once users start writing independently, allowing them to check their texts and catch spelling mistakes.

Last but not least, the ZLS is once again inviting the public to submit suggestions for the Wuert vum Joer, the Luxembourgish word of the year. Suggestions can be submitted until 30 November, and participants have a chance to win a prize. The ZLS will then draw up a shortlist, from which a jury will select the 2026 word of the year.

It is safe to say that it has been quite a year for the ZLS, and we have no plans to slow down in 2027!

Jeux de données 6