Test VocalCode in your language.
Help improve recognition.
VocalCode ships 25 European languages as model-supported previews — the models cover them, but honest per-language product testing needs something we cannot fake: a native speaker. VocalCode is free and open source for everyone. If you would like to help, complete the optional 20-minute test below and send your findings.
No coverage or review is expected. With your permission we can credit you by name or handle as a tester for your language.
The test — about 20 minutes
Install VocalCode (free, with no account or activation), pick your language on first run, and let the model download. Then open any place you can type — an editor, a chat, an email draft.
Dictate five ordinary sentences you would naturally say in your language — a message to a friend, an email opener, a to-do note. You write the sentences; they should sound like you, not like a phrasebook.
Dictate five sentences that mix your language with English technical terms, the way you would talk to a coding agent — words like function, async, deploy, commit dropped into your language's grammar. This is the hard case we most need tested.
Hold the key and talk for about a minute about whatever you are working on. Watch how it handles sentence boundaries, punctuation, and the moment you pause to think.
What to send
Email support@vocalcode.app with the subject “Native test: [your language]” and, for each part:
- Your OS and device (e.g. Windows 11 laptop, M2 MacBook Air)
- What you said (as text) versus what VocalCode typed — paste both
- Punctuation quality: natural, tolerable, or wrong for your language
- How it treated names, acronyms, and the English tech terms in Part 2
- Anything that crashed, hung, or confused you
- Your verdict: usable daily · usable with edits · not usable yet
Screenshots are welcome but optional. We read reports and use your findings to move your language from “preview” to a tested, documented status.
Languages we need
English and Mandarin Chinese already have heavy daily use; reports for them are still welcome. The languages below are the ones where a native speaker's 20 minutes changes what we can honestly claim:
All 24 preview languages share one auto-detecting Parakeet model; choosing a label does not constrain decoding. What your test measures is the whole path — microphone, model, punctuation, and text landing where your cursor is — in your language, on real hardware.
Why this is the honest way around
We deliberately publish no quantitative accuracy claims without evidence that would survive scrutiny. A native speaker saying “usable daily” or “not usable yet” about their own language, on their own machine, is that evidence — positive and negative reports are equally useful.