Does speech recognition understand Swiss German?

Speech recognition understands Swiss German only in part. It can produce usable text when the output you want is Standard German. It makes many more errors when you want the dialect itself, and no major platform we checked lists Swiss German as a language of its own. Expect to correct a lot of text, and test it on your own speech before you rely on it.

Swiss German is a family of dialects, and the public data covers all of them

Swiss German is not one language with one spelling. It is a family of dialects spoken across Switzerland. The STT4SG-350 corpus holds 343 hours of Swiss German speech from all dialect regions, and it is the largest public speech corpus for Swiss German ¹. Because the data covers many dialects, a model trained on it can handle more of the variation, but a single setting on a phone or computer cannot match every dialect.

Whisper gets about one word in four wrong on Swiss German without training

Dolev and colleagues tested Whisper on Swiss German audio in 2024. The large-v3 model, used without any training on Swiss German, had a word error rate of 23.0 percent on the STT4SG-350 test set ². That means roughly one word in four was wrong.

A Wav2Vec2 XLS-R model that the corpus authors fine-tuned on their own data reached 14.0 percent on its test set ³. Fine-tuning on Swiss German data cuts the error rate, which tells you that a general model is not tuned to the dialects.

The authors conclude that Whisper can be used for transcribing Swiss German without adaptation, provided the desired output is Standard German ⁴. Read that carefully: the conclusion is about Standard German text, not about writing in dialect spelling.

Platforms list a Swiss German region, not a dialect

Apple's on-device Dictation list for the Mac includes German (Switzerland) ⁵. The entry names a region, not a dialect, so it tells you which German the system expects, not that it recognizes Swiss dialects.

Windows 11 voice typing lists German as one language with no regional variant, and it has no Swiss German entry ⁶. Whisper also has no separate Swiss German code. It uses "de" for German ⁷, so an app built on Whisper will not offer a Swiss German setting by default.

A 2022 comparison favors a model built for Swiss German

In a 2022 evaluation, the FHNW team compared four anonymized commercial systems with its own model. The authors reported that their model had almost half the word error rate of the best commercial system ⁸. The FHNW team built that model, so read the margin as their own measurement, not an independent test.

For you, the practical point is that dialect-specific models exist in research, but they are not the default in the consumer tools most people use.

What to do

1. Decide whether you need Standard German or dialect text. If Standard German is enough, the Whisper result above applies, and you should expect correction of the words the model mishears.
2. Choose the German (Switzerland) setting where your platform offers it. Windows has only one German entry.
3. Plan for heavy editing if you need dialect spelling. The sources here do not show that the tools write dialect spelling reliably, so budget time for correction.
4. Test it on your own speech. Dictate a paragraph in the dialect you speak, count the errors, and compare with the 23.0 and 14.0 percent figures above. Your results will depend on your dialect and microphone.

The guide to speech to text in European languages compares German, French, Italian and Dutch lists, and the overview of languages and accents covers how accents and dialects affect recognition more broadly. For a related regional case, read the question on Brazilian and European Portuguese dictation. For a glossary entry on the term, see automatic speech recognition. The question on voice typing in Russian covers another language with its own grammar challenges.