Getting Swiss German Vocabulary Instead of Standard German

Most requests for “Swiss German” are actually requests for Swiss Standard German — ordinary written German with Swiss orthography and vocabulary. That is achievable and the rules are finite. Swiss German dialect is a different thing entirely, and mostly it is not what anyone shipping a product wants.

Two different things called Swiss German

German-speaking Switzerland is diglossic in much the same way the Arab world is. Swiss Standard German (Schweizer Hochdeutsch) is the written language: newspapers, laws, contracts, product documentation, signage. It is mutually intelligible with German German in writing and differs in a bounded set of spelling and vocabulary conventions. Swiss German (Schwiizerdütsch) is the family of spoken Alemannic dialects, used in almost all everyday speech and in informal writing such as text messages, with no standard orthography and substantial variation between cantons.

If your requirement is “a Swiss reader should not notice anything foreign”, you want Swiss Standard German, and the rest of this page is mostly about that. If your requirement is “dialogue that sounds like a person from Zurich”, that is the dialect, and the last section says what to expect.

A model asked for “German” produces German German, because Germany is roughly ten times Switzerland’s German-speaking population and supplies a corresponding share of the corpus. The same effect and the same fix apply to Austrian German.

Orthography: no eszett, and what follows

The single hard rule is that Swiss Standard German does not use ß. It is always written ss: Strasse, gross, Fussball, heissen, weiss, Grüsse. The Swiss Federal Chancellery’s writing rules make this binding for federal texts, and it holds across Swiss publishing generally. The Bundeskanzlei publishes its Schreibweisungen.

This is a genuinely useful property, because it makes the check trivial: any ß anywhere in the output is wrong for Switzerland. It is also the one that naive post-processing gets subtly wrong. Replacing ß with ss is safe in this direction, but the reverse is not — German German distinguishes Masse (mass) from Maße (measurements) and in Massen (in moderation) from in Maßen, so a Swiss text converted back to German orthography by a blind substitution introduces real ambiguity. Keep the Swiss text as the source if Switzerland is your primary market, not as a derivative. The normalisation issues around this are covered in German umlaut and eszett normalisation.

Two smaller conventions follow the same pattern: Swiss usage prefers French-style guillemets « » for quotation in many publications where Germany uses the low-high pair, and it is more tolerant of French loanword spellings retained unchanged.

Helvetisms

Vocabulary particular to Switzerland, most of it in everyday domains and much of it borrowed from French:

concept          Swiss (de-CH)      Germany (de-DE)
--------------------------------------------------------
bicycle          Velo               Fahrrad
pavement         Trottoir           Bürgersteig
ticket           Billett            Fahrkarte
platform         Perron             Bahnsteig
hairdresser      Coiffeur           Friseur
driving licence  Führerausweis      Führerschein
cream            Rahm               Sahne
chicken          Poulet             Hähnchen
pepper (veg)     Peperoni           Paprika
to park          parkieren          parken
to grill         grillieren         grillen
to phone         telefonieren       anrufen
mobile phone     Handy / Natel      Handy
rubbish          Kehricht / Abfall  Müll
apartment        Wohnung            Wohnung
hello (informal) Grüezi             Hallo / Guten Tag

Note the verbs. Parkieren and grillieren are the kind of difference a vocabulary list built from nouns misses entirely, and they are common enough to be noticed. Natel is generational — it was the Swisscom brand name and is now dated, with younger speakers using Handy or Mobiltelefon; putting it in output aimed at a young audience reads as if written by someone’s uncle.

Numbers, dates and money

These are locale data rather than vocabulary, and they are the details that make a page look Swiss or not regardless of the words:

• Thousands separator. Swiss convention uses an apostrophe: 1’000’000. Germany uses a full stop: 1.000.000. This is the most visible single difference on any page with figures on it.
• Currency. Swiss francs, written CHF before the amount, and amounts are conventionally written with the decimal part as CHF 12.50 using a point, not a comma — unlike the German-language decimal comma used elsewhere. Financial contexts also use the .– form for whole francs (CHF 12.–).
• Dates. DD.MM.YYYY as in Germany, so no change — but see day-month and month-day formats for why this is worth stating explicitly anyway.
• Phone numbers. Country code +41, and the national trunk zero is dropped in international form.

The practical instruction is to ask for “Swiss Standard German (de-CH)”, state the ss rule explicitly, list the ten Helvetisms relevant to your domain, and specify the apostrophe thousands separator. Then grep for ß and for the German column above. That combination catches nearly everything a Swiss reader would flag.

Switzerland is not one locale

A product that says it supports Switzerland with a single de-CH string set has covered roughly two thirds of the country. Switzerland has four national languages — German, French, Italian and Romansh — and the first three are all used in official and commercial contexts, with the relevant locale tags de-CH, fr-CH, it-CH and rm-CH.

Swiss French is worth singling out, because it contains the most famous locale difference in the language and one a model will get wrong every time unless told. Swiss French uses septante for seventy and nonante for ninety, where French from France uses soixante-dix and quatre-vingt-dix; parts of Switzerland, including Vaud and Geneva, additionally use huitante for eighty against quatre-vingts. Belgian French shares septante and nonante but not huitante. Any generated French containing a number between seventy and ninety-nine is therefore variety-marked, whether you intended it or not, and there is no way to write it neutrally.

Swiss French also shares the Helvetism pattern with Swiss German — natel, déjeuner for the midday meal, and a set of administrative terms specific to the cantonal system — so the same technique applies: name the locale, supply the vocabulary, check mechanically. The parallel with Quebec French is exact, except that Switzerland has no terminology authority of its own to point a model at.

Romansh, with roughly 40,000 speakers and five written idioms plus the constructed standard Rumantsch Grischun, is a genuinely low-resource language and should be treated as one. Do not assume a model can produce it usefully; assume it will produce something that looks like it, and that the five idioms will be blended together in the output because nothing in the training data separates them.

If you actually want the dialect

Swiss German dialect in written form is a genuinely low-resource target, and it is worth being clear about why. There is no standard orthography, so the same word is spelled several ways by different writers. The dialects differ substantially between Zurich, Bern, Basel and Valais, so “Swiss German” does not name a single target. And the written corpus is thin and informal — chat messages, some social media, occasional literary work — because anything formal is written in Standard German.

Expect a model to produce something that gestures at the dialect rather than something a speaker would accept: Standard German with a few dialect words substituted, inconsistent spelling, and grammar that stays standard. If you need it, name the specific canton, supply several sentences of example text in the exact orthography you want, and have a speaker review the result. This is the general failure pattern described in what a model does with a language it barely knows, applied to a variety whose speakers mostly do not write it down.