Language machines: translation and Swiss German

Read the PDF edition (2 pages)

Part J showed that Basel is a multilingual city in which Standard German, Baseldytsch, French and English meet every day. Language technology is therefore of particular importance for its institutions. Neural machine translation, which Part J, Chapter IV discussed in relation to contract drafting, has become a routine tool in banks, insurers and pharmaceutical groups. It is used to understand documents in languages that staff do not read, to produce first drafts of client correspondence and to check the consistency of several language versions.¹ Large language models have extended these capabilities, because they can translate with attention to context and terminology and can be instructed to follow a glossary.

Speech recognition, the second field, poses particular problems in German-speaking Switzerland. Part J, Chapter I described how the absence of a standard spelling and the diversity of dialects made Swiss German difficult for systems trained on Standard German. Research groups at Swiss universities and universities of applied sciences have built corpora of Swiss German speech. Since 2022, systems that transcribe dialect into Standard German text have improved considerably.² For Basel's banks and insurers, which serve clients by telephone and increasingly through voice assistants, this improvement has practical consequences. Calls can be transcribed and summarised, the client's request can be routed automatically and the records required under financial services law can be prepared with less effort.

The use of speech technology in client relationships raises its own legal questions. The recording of telephone conversations requires the consent of the participants under Article 179bis and following of the Criminal Code, unless the law provides otherwise, while securities dealers have specific recording duties under financial market law. Voice biometrics, which some institutions use to authenticate callers, process biometric data, which the Federal Act on Data Protection treats as sensitive personal data.³ The rise of voice cloning, discussed below, has also made voice authentication less reliable than it once appeared.

Language technology may change the position of languages in the financial centre in a subtler way. If machines translate reliably between German, French and English, the pressure to use English as the single working language, which Part J, Chapter III described, could ease. Staff could write and speak in their own language while colleagues and clients read and hear it in theirs. The opposite effect is equally plausible, since models trained predominantly on English text may perform best in English and reinforce its dominance.⁴ Which of these effects prevails will depend on the quality of models in German, French and Swiss German, an area where Swiss research institutions have a direct interest.

Notes

1. See Part J, Chapter IV, section on machine translation, AI and professional secrecy. To be verified. ↩
2. See Part J, Chapter I, section on school policy, media and digital communication; on Swiss German speech corpora, such as the Swiss Parliaments Corpus and STT4SG-350, see the proceedings of the SwissText conferences. Status to be verified. ↩
3. Arts. 179bis to 179quater of the Swiss Criminal Code; Art. 5(c)(4) FADP on biometric data uniquely identifying a natural person as sensitive data. To be verified. ↩
4. On the performance of language models across languages and the predominance of English in training data, see the technical reports of the model providers and the research on multilingual language models since 2023. To be verified. ↩

For general information only. This article does not constitute legal or tax advice, does not replace advice on an individual case and creates no lawyer-client relationship. It reflects the legal position as of August 2025.