Prototype by Sandro Barbieri, University of Zurich, CAS Generative AI, Class Fall 2026.

About Clerina

Romansh has five regional idioms – Sursilvan, Sutsilvan, Surmiran, Puter and Vallader – and a common written standard, Rumantsch Grischun. The app shows visitors the eleven everyday phrases (and the numbers 1–10) in the Romansh idiom of the village they are in, so they can greet locals in their own language.

Romansh sources and partners

Clerina is built on the work of the Romansh language community and research. Thank you to:

Lia Rumantscha: Lia Rumantscha – the umbrella organisation of the Romansh language; behind Pledari Grond and ALAS and co-developer of the ZurichNLP translation model.
Pledari Grond: Pledari Grond – the Romansh dictionaries of the Lia Rumantscha (data CC BY 4.0). Source of the phrases and numbers for Rumantsch Grischun, Sursilvan, Sutsilvan and Surmiran.
ALAS: ALAS – the translator of the Lia Rumantscha, used to cross-check translations.
Uniun dals Grischs: Uniun dals Grischs – publisher of the Dicziunari for Puter and Vallader, the reference for these two idioms (not yet checked for Clerina); also contributed dictionary data to the translation model.
ZurichNLP, University of Zurich: Romansh NLLB-200 1.3B translation model – Jannis Vamvas, Ignacio Pérez Prat, Angela Heldstab, Dominic P. Fischer, Sina Ahmadi, Rico Sennrich, with Lia Rumantscha (CC BY-NC 4.0); used to propose Puter and Vallader phrases. Text collections Romansh theatre plays and La Quotidiana, used to check which forms people actually use.
La Quotidiana: La Quotidiana – the Romansh daily newspaper; its articles (via ZurichNLP) show how the idioms are written today.
RTR: RTR – Radiotelevisiun Svizra Rumantscha – contributed texts, translations and transcriptions used to train the translation model.
Canton of Graubünden: Canton of Graubünden – Romansh press releases used to train the translation model.
Mediomatix: Mediomatix – Romansh school texts in all idioms (ZurichNLP), used to train the translation model.

How this app was built

The idea
Romansh has five regional written forms (idioms) plus a common standard. A greeting that sounds right in Scuol is different in Disentis. Clerina shows the phrases that fit the place you are in.
Finding the right words
An AI translation model from the University of Zurich suggested phrases. Because AI can be wrong – it once offered “Bunura!” (early morning) as a greeting – every phrase was then checked against the official dictionary of the Lia Rumantscha, and thousands of real Romansh texts were searched to see which form people actually use.
Adding sound
A computer voice read each phrase aloud. A second AI that understands spoken Romansh listened to every clip; only clips it understood correctly were kept. They are marked as computer voice, ready to be replaced by recordings of local speakers.
Knowing where you are
Your browser can share its location. Clerina compares it with a list of Romansh villages and picks the local idiom. You can also type a town or postcode; the swisstopo place-name search finds it.
Being honest about limits
Anything not confirmed by the dictionary is marked with †. Puter and Vallader still need a check by native speakers.

Main parts

Everything above the line was prepared in advance and checked; the running app only shows that prepared, checked content.

Technologies

Computer voice: OmniVoice – k2-fsa (Zhu et al., 2026). Model CC BY-NC; audio tokenizer under the Boson Higgs Audio 2 Community License.
Pronunciation check: whisper-medium-romansh – Romansh speech recognition by Hugging Face user manueltuor, based on OpenAI Whisper.
Map: National map (Landeskarte) and municipality boundaries (swissBOUNDARIES3D) © swisstopo, Federal Office of Topography – free use under the Swiss open government data terms. Map display with Leaflet (BSD-2-Clause). Idiom areas are the municipalities of the villages listed in Clerina and are approximate.
Town lookup: Place names and postcodes from the search service of geo.admin.ch © swisstopo – free use under the Swiss open government data terms.
App and hosting: Gradio, Hugging Face Spaces, CTranslate2, Hugging Face Transformers, PyTorch, Python.
Design: Font Inter (SIL Open Font License). Colours inspired by Graubünden's autumn hiking scenery. No official logos or photos are used; not affiliated with any tourism organisation.
AI assistant: Planned and built in conversation with Claude Code (Anthropic): code, design, data checks and this page.

Imprint

Clerina is a non-commercial study prototype by Sandro Barbieri, created for the CAS Generative AI at the University of Zurich (Class Fall 2026). Translations and computer voices may contain errors; this is not an official language resource. Source code: Hugging Face Space.