Blog Engineering 3 min read
Transcribe meetings in 25 European languages on a Mac, on-device
Everpaper transcribes meetings on your Mac with NVIDIA's Parakeet TDT 0.6B v3, which covers 25 European languages and detects the language automatically. Nothing is uploaded and no bot joins the call. NVIDIA's model card reports an average word error rate of about 8 to 12 percent across its multilingual benchmarks, so results vary by language and audio.
Disclosure. Everpaper is our product, and it is in this article. We say where it falls short, and every competitor fact below carries the date we checked it.
Facts as of October 2026. Prices and features change. Each source is linked at the end.
Most meeting tools were built for English first. If your calls are in German, Italian, Polish or a mix, the useful questions are which languages are covered, whether the tool picks the language itself, and whether the audio has to leave your machine.
The model
Everpaper’s meeting transcription uses NVIDIA’s Parakeet speech model, specifically the multilingual v3 release with about 600 million parameters, published under CC-BY-4.0. It runs on the Mac. The model card lists 25 languages:
Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian.
It detects the language of the audio on its own. You do not pick one before the meeting.
What the accuracy numbers say
NVIDIA reports average word error rates for the multilingual set of about 11.97% on Fleurs, 7.83% on MLS and 11.98% on CoVoST, measured with greedy decoding, without an external language model, and after removing punctuation and capitalization. On the English Open ASR Leaderboard average it reports about 6.3%.
These are NVIDIA’s benchmark numbers. They are averages over many languages, and individual languages score better or worse. We have not published our own per-language measurements, so test a sample of your own audio before you depend on it. Noisy rooms, crosstalk and accents move the result more than the model version does.
How it works in a meeting
Press ⇧⌘M from any app. Everpaper records your microphone and the call’s system audio as separate time-aligned tracks, transcribes on the Mac with live speaker labels, and saves the meeting to your Library. The default Everyday mode is fast and gives live captions. Detailed mode re-listens to the whole call after you stop, needs a one-time download of about 2 GB, and takes longer. You can re-run one meeting in either mode later.
Nothing joins the call and recordings are never uploaded. Because no bot is in the room, tell the other people you are recording. A saved transcript exports as text, Markdown, SRT, WebVTT or JSON, and the audio saves as an .m4a file. You can also import a WebVTT transcript from another tool.
Summaries need an AI provider you set up. An on-device provider keeps them on the Mac, and a cloud provider receives the transcript only after you turn on a switch that is off by default.
How others compare on languages
| Tool | Languages listed on its pricing page |
|---|---|
| Everpaper | 25 European languages by default, automatic detection; Whisper optional in Settings |
| Otter | English, Spanish, French, German, Japanese, Chinese |
| Fireflies | English, Spanish, French, Portuguese, Italian; 100+ more in beta |
| Granola | Multi-language support on all plans, no list given |
Parakeet v3, the default, does not cover Japanese, Chinese, Arabic, Hindi or Turkish. For those, Settings, Speech also lets you switch the speech engine to Whisper Large v3 Turbo, which the app describes as 99 languages with automatic detection (a separate model download; we have not measured it per language). Otter lists Japanese and Chinese on its pricing page.
Limits
Meeting transcription uses Parakeet v3 and the 25 languages above unless you switch the speech engine to Whisper in Settings. Dictate, which types what you say at the cursor, works in Italian and English only for now. Everpaper needs macOS 15 or later on Apple silicon, has no mobile app, and is in private beta with no public price.
More on choosing a tool: best AI meeting assistants.
Apply for the Everpaper private beta.
Frequently asked questions
Which languages does Parakeet v3 support?
Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian.
Does it detect the language automatically?
Yes. NVIDIA's model card says the model automatically detects the language of the audio and transcribes it without additional prompting.
How accurate is it?
NVIDIA reports multilingual average word error rates of about 11.97% on Fleurs, 7.83% on MLS and 11.98% on CoVoST, using greedy decoding without an external language model and after removing punctuation and capitalization. Those are benchmark averages, not a promise for your recordings.
Does the audio leave my Mac?
No. Everpaper records and transcribes meetings on the Mac and never uploads recordings. Only text can reach an AI provider, and only one you set up.
Can I dictate in all 25 languages?
Not yet. Everpaper Dictate currently works in Italian and English. The 25-language coverage is for meeting transcription.
Sources
Written by Marco Iannello Updated 21 October 2026 How we write comparisons