"How many languages do you support?" — and why our honest answer is six numbers, not one

It is the first question on every procurement call and the first question in every demo. We get it three times a week:

"How many languages does InterMIND support?"

The honest answer is: it depends on which surface you mean. We have six of them, and they have different language lists for good reasons. This page is the definitive answer so we stop giving inconsistent ones.

The short version

The product is 24 languages. Files: 30. Site: 17. API engine: 24.

Voice, chat, and notes share the same 24 — the full engine list. Files are wider because the file pipeline can be. The website is narrower because translating a marketing site is a different cost than translating in-product copy.

Everything below is why those numbers are not the same.

Why there isn't a single number

A translation platform is not one product. It is at least three:

1. A real-time speech pipeline — audio in, translated audio out under one second, in a meeting where the latency budget is brutal.
2. A real-time text pipeline — chat messages and shared notes, character-by-character, propagated to every viewer in their language.
3. An asynchronous document pipeline — a 40-page PDF dropped into chat, translated as a whole file with formatting and structure preserved.

Each pipeline has its own engine, its own latency budget, and its own per-pair quality envelope. A language that is production-grade for documents (slow, careful, many passes) can be unusable for live voice (fast, one shot, no retries). Treating those as "one language count" is what produces marketing pages that say "200+ languages" and tell you nothing.

So we publish per-surface.

1. Platform API — 24 languages

This is the widest list, because it is the engine surface. The createParticipant endpoint accepts these 24 codes:

ar, cs, da, de, en, es, fi, fr, hi, hu, is, it, ja, ko, nl, no, pl, pt, ro, ru, sv, tr, uk, zh

If you are an integrator building on Mind directly, you can hand any of those 24 to a participant. That does not mean every one of those 24 scores the same on our quality bar. It means our engine can emit text in that language. The product shows all 24 and publishes each pair's quality openly (next section).

2. Voice, chat, and notes inside the product — 24 languages

When a user joins an InterMIND meeting, the language picker shows all 24 engine languages — including Arabic (ar) and Hindi (hi), which we previously held back.

We used to remove those two, because some of their pairs don't clear the median-quality bar we publish on /benchmark. We've changed the trade-off: instead of hiding a language wholesale, we surface it and let the public per-pair numbers tell the truth. Every pair — strong or weak — is measured on real traffic and deep-linkable by URL, so a user (or an auditor) can check en→ar before they rely on it rather than discover its quality mid-meeting.

This is still the honest version, just tuned. Most of the category lists everything the model can emit and stays silent on quality; the opposite extreme (hide anything below a bar) makes a real, in-demand language invisible. We do neither: show the full list, publish the number next to it. If a pair isn't good enough for your use case, the benchmark says so out loud.

The same 24-language list applies across the three real-time surfaces:

• Voice translation — per-viewer translated audio with sub-second latency. Each participant picks their own listening language at the start of the call.
• Chat messages — every message translated as it is typed; edits produce per-language diffs (see v1.2).
• Shared notes — character-by-character live translation of the host's notes pane, per viewer, with diff history.

One picker, one set of 24 languages, three places it shows up.

3. On-demand file translation — 30 languages

Drop a PDF, DOCX, PPTX, or XLSX into the chat. Each participant can request the file in their language. The translated copy is returned as the same file format with structure preserved (tables, headings, lists).

This surface uses the DeepL Document API; our file pipeline maps 30 target languages — wider than our real-time voice pipeline. If you can translate the PDF to Estonian on DeepL today, you can translate it to Estonian in InterMIND chat today.

The file list includes some languages our real-time pipeline does not — for example Bulgarian, Greek, Estonian, Indonesian, Lithuanian, Latvian, Slovak, Slovenian. Both Arabic and Hindi are now on the real-time voice/chat/notes list as well as, for Arabic, the file list; a French participant can both listen to the meeting in Arabic and request the contract PDF in Arabic.

The one real-time language not on the file list: Hindi. Our document pipeline doesn't map it yet, so Hindi is available for live voice, chat, and notes but not for on-demand file translation — the reverse of the old Arabic asymmetry. We flag this in the file picker rather than hide it.

Why the file list is bigger in general:

• Documents are asynchronous. There is no one-second budget, so the pipeline can afford a slower, more careful engine that handles more pairs well.
• DeepL is a dedicated document-translation engine, and we use it directly for files. We do not try to route documents through the same engine that runs voice. Different problem, different tool.

4. Website UI — 17 languages

intermind.com is currently shipped in 17 locales: English, German, Spanish (Latin America), French, Italian, Portuguese (Brazilian), Dutch, Polish, Ukrainian, Chinese (Simplified), Russian, Japanese, Korean, Turkish, Arabic, Indonesian, and Vietnamese.

Two of those — Indonesian and Vietnamese — are localized on the public site but are not yet available for in-meeting translation: our speech engine doesn't cover them. So you can read the landing pages, blog, and docs in Indonesian, but if you join a meeting, voice and chat translation won't run in Indonesian yet. We flag this in the language switcher rather than let you discover it in a live call. The other fourteen locales are full products — site and in-meeting both.

Why narrower than the file list:

• A marketing site has its own translation cost — and its own quality bar. Landing pages, FAQs, pricing, legal copy. It is not free.
• We ship locales where traffic and pipeline justify the maintenance, not for every language the product supports.
• The site list leads the product in two places (id, vi) on purpose: localizing the public surface is a standalone capture layer even before the speech engine reaches a language.

This list will grow as pipeline grows. It is not capped on principle. It is capped on cost.

The pattern, and how to read it on any vendor page

Every multilingual vendor has these same surfaces. Most of them quote you the widest one and let you discover the others on your own. The honest version is:

• The engine list is widest, because it is what the model emits.
• The product list is narrower, because not every engine output meets a product bar.
• The document list often differs from the voice list, because the engines are different.
• The website list is narrowest, because translating your own site costs you cash.

When you evaluate any platform in this category, ask for the four numbers separately. If a vendor refuses to split them — or doesn't know — you are not getting a useful answer.

What we will not pretend

• Some pairs are weaker than others, and we say so. Arabic and Hindi are back in the voice/chat/notes picker, but not every pair scores the same. Rather than hide a whole language, we publish per-pair quality on /benchmark — a weak pair like en→ar shows its real number instead of being quietly dropped. Check the pair you need before you rely on it.
• The file list (30) being wider than the voice list (24) is a real gap, not a feature. A French user can request an Estonian translation of the PDF but cannot listen to the meeting in Estonian. We will not paper over this by quoting the bigger number and hoping you don't notice.
• The site leading the product in Indonesian and Vietnamese is a gap, not a feature. You can read the site in those two but not yet run a meeting in them. We flag it in the switcher instead of letting you find out mid-call. We close it when the speech engine adds them.

Try it yourself

• Try the live demo — runs the production voice + chat pipeline against your audio, in any of the 24 product languages. The same pipeline that scores /benchmark.
• See the benchmark — per-pair, per-month quality on real traffic. Every pair in the picker, strong or weak, deep-linkable by URL.
• Read the methodology — what the numbers are, what they aren't, who the judge is.

Six numbers, six surfaces, one engine. That is the honest answer to "how many languages do you support."

FAQ

How many languages does InterMIND translate live?

24 — the same list across voice, chat, and shared notes: ar, cs, da, de, en, es, fi, fr, hi, hu, is, it, ja, ko, nl, no, pl, pt, ro, ru, sv, tr, uk, zh. Every pair's quality is published per month at /benchmark.

How many languages does document translation support?

30, via the DeepL Document API — drop a PDF, DOCX, PPTX, or XLSX into chat and request it in your language with formatting preserved. The file list is wider than the live list (it adds e.g. Bulgarian, Greek, Estonian) but misses Hindi.

Why is the website in fewer languages than the product?

The site ships in 17 locales because localizing a marketing site is a separate cost from in-product translation. Two site locales (Indonesian, Vietnamese) lead the product on purpose — flagged in the language switcher.

— The Mind.com Team

Sources: DeepL — supported languages, Mind API — createParticipant; product surface lists verified against the shipped code, checked August 2026.