How a Swiss start-up makes DeepL and Google look pretty old

Textshuttle launched a free AI-based translation service this week, declaring war on DeepL and Google Translate.

The young team emphasizes that the software has been developed entirely in Switzerland and runs on local servers. You can use it to translate large amounts of text, including Swiss German and Romansh.

Watson asked those responsible how they came up with the crazy idea of ​​developing a translation app for the numerous Swiss-German dialects, how they protect the sensitive data entered by users and whether their AI is also prone to hallucinations.

Textshuttle co-founder and CTO, Samuel Läubli, and CEO Lucas Seiler answered the questions together.

The interview

How did you come up with the crazy idea of ​​developing your own voting robot and using it to challenge big tech companies?
Because we were frustrated. Martin and Samuel, who built Textshuttle, are doing basic research on AI speech processing. In the academic field, it became clear as early as 2015 – long before DeepL – that the use of large neural networks would represent a qualitative leap forward for automatic text translation.

The fact that at that time this technology was only made available to the general public by companies abroad, who were not interested in “minor” languages ​​such as Rhaeto-Romance, linguistic idiosyncrasies such as “street” instead of “street” or the corporate phrasing from a Swiss retail company, urged us to start developing our own speech robots with the best minds in AI speech processing at the Zurich site – and to this day without external funding.

Why are you taking the step now to offer the translation tool for free?
Simply put: Switzerland deserves a platform that speaks our national languages ​​and processes content in Switzerland. In 2023, no one would have to write texts without AI support.

It is of course very smart and entrepreneurial to take advantage of the current AI hype. How do you see the current discussions about the risks and dangers of generative AI?
We are annoyed that every second piece of software nowadays has an AI label. Depending on one’s point of view, that’s not a bad thing either: the first AI systems were rule-based, so the term can be used relatively broadly, depending on how the term is interpreted. We prefer to talk about machine learning or deep learning, i.e. the use of (now very large) neural networks, and with Textshuttle we have done nothing else from the start. When we say “AI”, we mean “real, modern machine learning” – and you can see that when you compare our tool with other systems, such as Google Translate or DeepL.

Incidentally, we do not see AI as a hype, even though a lot of marketing is done with it. After the digital transformation from 2000 to 2020, we are at the latest with the arrival of LLMs, etc., at the beginning of the next disruption, which is rightly called “AI transformation”. It’s not just slobs who say that. Even reasonably sober heads such as the German Minister of Labor Hubertus Heil are now convinced that from 2035 there will be no more jobs that have nothing to do with AI.

What about the risks?
Generative AI certainly has the potential for risk and danger; at the same time it offers enormous opportunities. We take the position that the first thing to happen is a serious discussion of the subject at all levels. This is especially true for educational institutions, but every competitive company will also need an AI strategy in 2023. It’s exciting to see that companies like Ringier have not only recognized this, but made it a top priority.

I have an additional question (with a wink): Can Textshuttle also hallucinate?
We’ve been able to greatly reduce hallucinations in our models over the past few years, so the problem is hardly significant – perhaps our system will “hallucinate” a punctuation mark in a translation where there is none in the source text. On rare occasions, we – and all of our AI translation competitors – experience the opposite: omissions, i.e. information in the source text that is omitted in the translation.

Key word: security

What about data protection at Textshuttle?
Our clients include renowned Swiss banks and insurance companies that entrust us with their sometimes very sensitive texts. This shows that we take data protection very seriously. The translation takes place on our own servers in Switzerland. This gives us full control over the entire production process.

How is it ensured that sensitive data entered by users does not fall into the wrong hands?
We specialize in handling sensitive data and implement a large number of technical and organizational security measures. They also meet the high demands of our customers – especially from the financial sector. We are regularly checked by them.

As a very privacy-conscious user, what exactly can I imagine this means?
To prevent data from falling into the wrong hands, one of the most important and effective organizational measures is to keep the group of people with access to data as small as possible (“need-to-know principle”). In addition, we carefully select employees and make them aware of privacy issues in the development process.

And of course we also have a wide range of technical measures in place: from biometric access controls to our servers, firewalls to transfer controls that ensure that data cannot be read by unauthorized persons, just to name a few. We have to report this regularly to our customers.

Why is Textshuttle so good at Swiss German, while other providers seem to struggle?
Big companies focus on big languages. This is relatively easy because AI methods for automated language processing are data hungry. In other words, if I show my AI 100 million sentences that people have translated from English to German, it can easily learn to imitate them – it doesn’t need to generalize as much because it sees an example for almost every word in almost every word. context.

It is different with languages ​​such as Rhaeto-Romance or Swiss: there are probably barely 100 million sentences in the world that have ever been translated from English into Bernese German. Not to mention that residents of Bern and Emmental would write the same sentence differently.

In the field of machine learning, this is an exciting challenge for us. Based on fundamental research at the University of Zurich, we have developed methods that allow neural networks – on which modern AI applications are based – to learn much more from much fewer human-made translations. But that’s just the beginning: our AI gets better, especially for Swiss dialects, the more people use it on our platform.

Will Textshuttle ever master Wallisertiitsch?!
Technically this is absolutely feasible! We’re answering the question – you’re welcome to do a survey of the watson users, what dialect do they want next. 😉

Regarding the financing: Do I understand correctly that there is cross-subsidization through the paid enterprise solution?
On the contrary, it is an investment in our business solution! Now we have the opportunity to make new features (such as gender-fair language) available to the public and optimize them to meet the professional requirements of a translator at a Swiss insurance company.

What can you say about Textshuttle’s power consumption?
Our language models are significantly faster and more energy efficient than the models behind ChatGPT, for example, because they are specialized in translation tasks. However, Textshuttle’s areas of application will expand rapidly, for example when it comes to editing texts in the same language. Example: AI converts a complicated newspaper article into a plain-language version or into gender-sensitive German.

What do you mean?
Large neural networks (e.g. LLMs) can have different “sizes” – we define and optimize the optimal number of parameters. If we expand the offer on our platform in the future with applications beyond classical translation, the networks and thus the power consumption may also increase. However, it is to be expected that this can be compensated or at least absorbed by more efficient hardware.

Hand on heart: how much electricity is consumed?
The most energy-intensive part of our processing chain is the machines on which the actual translation takes place. At our current scale, these machines require about 4200 kWh of electricity per month at full capacity. They are stationed in Switzerland and run on electricity from 100 percent renewable energy (mainly hydropower). We should be able to cover Switzerland’s translation needs even if a large number of people will use our platform one day.

Textshuttle is said to “allow you to translate larger amounts of text and more documents than other free offerings.” Specific?
Registered users can translate any number of texts up to 15,000 characters long (compared to 5,000 with DeepL) on textshuttle.com. In addition, they can translate three documents per day, with DeepL only three documents per month are possible. With Textshuttle, translated documents are editable and ad-free.