Why supporting 80+ languages is harder than it sounds

In a recent customer conversation, I was told: "Translation is just a commodity feature these days." Reflecting on what seemed like a provocation at the time, it certainly feels true with translation options popping up in many applications and service. But also true is that translation quality is not equal across languages!

Mostly that's not because some languages are inherently "harder," but because AI models learn from data, and the overwhelming majority of usable speech and text data covers a few dozen languages, much of it paired with English. Supporting 80+ languages at high quality means dealing with data scarcity, script diversity, dialect variation and code-switching, all at telephone audio standard quality.

When we announced 80+ languages, the most common (and best) skeptical question was: are they all equally good? Honest answer: no system on earth achieves that today. Here's why, and how we deal with it.

The data pyramid: high-resource vs. low-resource languages

English, Mandarin, Spanish and a handful of others sit atop mountains of transcribed speech and parallel text. A step down, languages like Polish or Thai have solid but thinner resources. Further down are the so-called low-resource languages: Amharic, Hausa, Khmer and others spoken by tens of millions of people, yet with startlingly little machine-usable data.

Multilingual speech recognition and translation models like Whisper and SeamlessM4T have pushed coverage enormously through transfer learning: knowledge from data-rich languages genuinely helps data-poor relatives. But transfer has limits, and they show at exactly the wrong moment: proper names, numbers, negations.

It's not just data volume

Those are my top three additional aspects I want to point to, when talking about complexity:

Scripts and orthography: some languages lack standardized spelling, write vowels optionally (Arabic), or have no word boundaries in text (Thai). Each quirk complicates the handoff from recognition to translation.

Dialect continua: "Arabic" on paper is Modern Standard Arabic. On a phone call it's Egyptian, Levantine or Maghrebi, which are nearly different languages acoustically.

Code-switching: real bilingual speakers mix languages mid-sentence ("Ich hab dem landlord gesagt…", German for "I told the landlord…" with an English word dropped in). Monolingual pipelines handle this poorly, and real calls are full of it. Good luck, when you deal with Indians living in the large metropolitain areas like Mumbai and speaking a mix of many Indian languages wrapped into one. Real life is messy, which makes it challenging and beautiful somehow.

How a language makes it into our portfolio

There is no single best translation model! Most models learn primarily from data paired with English, and their published accuracy figures are usually measured against English too. But most of our customers don't translate into English. They translate between German and the language of the person on the other end of the line, so that's what we measure.

Every language is tested twice, always with German on one side. First, we test it technically, with automated benchmarks for recognition and translation quality between German and the target language. We run these on telephone-quality audio rather than clean studio recordings. Then native speakers judge what metrics can't: whether names, numbers and negations survive, whether the register fits a professional conversation, and whether the output would hold up in a real call. Metrics tell us where to look; native speakers tell us whether it works.

A language joins our portfolio only if it meets all four criteria. We evaluate several AI models for every language. A model is eligible only if it meets our quality bar for German and that language, and its provider is GDPR-compliant, does not train on your conversations, and does not store them (Zero Data Retention). Among the eligible models, we use the one that performed best for that language. If no model meets all four criteria, we don't offer the language, however often it's requested. When models improve or new ones appear, we test again, and that's how new languages are added.

That's why our language list is shorter than it could be. For businesses and public authorities, a mistranslated date, amount or deadline has real consequences. Every language we offer has to be one we can stand behind.

Announcing 80+ languages is easy. Supporting them honestly is the actual product: testing each one against German, by machines and native speakers, and leaving out any language that doesn't meet the bar. That's the work.

Planning a rollout? Book a discovery meeting.