A German Startup Races Google To Develop A Universal Translator

Experts say Cologne-based DeepL holds its own in using cutting-edge AI for translation services.

Kutylowski at DeepL’s offices in Cologne.
Kutylowski at DeepL’s offices in Cologne.

Techies have been promising for years to make a real-life version of a universal translator—the gadget that appears in myriad sci-fi books, movies and TV shows, allowing characters to understand one another regardless of the language they’re speaking. Rapid advances in artificial intelligence, which grabbed the public imagination with the 2022 introduction of ChatGPT, have made that imagined reality actually seem within reach.

One logical candidate to offer such a translator is Alphabet Inc.’s Google, whose Translate service already handles more than 130 languages. Last year, Google showed a video featuring eyeglasses that could display subtitled translations in real time. (The actual glasses have yet to arrive.) There’s also OpenAI, the maker of ChatGPT, which users have been employing to translate multiparagraph passages. Spotify Technology SA recently disclosed plans to use OpenAI’s tech to translate popular podcasts into other languages, read by AI-rendered clones of their popular hosts’ voices.

A less prominent candidate, but one that industry insiders have been tracking for years, is the 700-person startup DeepL SE in Cologne, Germany. This January it raised funds at a €1 billion ($1.1 billion) valuation. In December the company plans to introduce its first voice interpreter, a feature that automatically captures a speaker’s words, then translates and transcribes them into text in another language. Eventually, DeepL plans to bring this feature to its own app, as well as to other services such as Zoom. Jarek Kutylowski, its founder and chief executive officer, says he imagines such a translator living “within each and every business meeting,” making language barriers irrelevant.

Unlike its sprawling competitors, DeepL is solely focused on machine translation. As AI spreads across businesses, it’s uncertain whether a few general-use models will dominate the market, or if many organizations will flourish offering tools that excel at specific tasks. DeepL’s continued success would point to the latter possibility.

The startup, which began operating through a bare-bones website in 2017, now covers 31 languages and sells a paid version to more than 20,000 clients, including law firms and consulting companies, many of them in Asia. DeepL says “tens of millions” of people use its service every month, and it plans to open its first US office in January. Ajay Vashee, a partner with IVP, a venture capital firm that’s invested in DeepL, compares the startup to Dropbox and Slack, other household names in the software industry that his firm has invested in.

Intento Inc., which tracks machine translation, recently ranked DeepL with Google and Inc. for best overall offerings and placed DeepL higher in some sectors such as education, health care and finance. A recent academic study showed DeepL made fewer errors than Google Translate when going from English to Polish, though DeepL has also been found to have more trouble than ChatGPT in maintaining the gender of words during translations. An Amazon representative declined to comment, and Google and OpenAI didn’t respond.

The most advanced translation software today relies on neural networks, which are better than previous technologies at producing translations that accurately place words and phrases in context, says Eva Vanmassenhove, an assistant professor at Tilburg University in the Netherlands, who studies the topic. As an example she offers the word “bank,” which could be a financial institution or the earth beside a river; a sophisticated translation system needs to be able to choose the definition based on how it’s being used.

Kutylowski describes DeepL’s tech as a “contextual engine,” which works by absorbing entire paragraphs of text at once. He says DeepL is using OpenAI tech for some experiments while also developing its own large language model. DeepL’s system is a “grade smaller” than more general-use models, Kutylowski says, but also more efficient, because it’s tailor-made for translation. The company uses special web crawlers it’s built to locate translations online and assess their quality, according to a description on its website. It also pays thousands of contractors to offer human feedback to train its models.

Beyond this description, Kutylowski doesn’t offer many specifics, citing intellectual property concerns. This is common among tech companies, Vanmassenhove says. “It’s always difficult to know what’s going on and how they compare,” she says. She cautions that existing research is still fairly thin and that the technical methods from companies are constantly shifting. “I don’t dare to make any claims about DeepL being any better,” she adds.

Forecasters project the global market for translation services will top $44 billion in the next decade, with the market likely driven in part by multinational corporations and institutions looking for cheaper options than human translation. To win such business, Kutylowski knows he may have to outperform Google Translate, and he’s confident that his company is already proving it can do just that. “We’ve shown that, with a lot of focus and determination, we can keep up even with a competitor that is bigger,” he says.

More stories like this are available on

©2023 Bloomberg L.P.