Transcription that never leaves your computer.
Lunosia turns speech into text using a model that runs inside your own browser. No account. No upload. No server to send anything to — which is why it still works with your internet disconnected.
Free for microphone recording in 99 languages · $19 once for everything else · no subscription
Don't take our word for it. Check.
Every transcription tool claims to respect your privacy. The difference here is that you can verify it in about thirty seconds, without trusting anyone — including us.
Load the app once. It downloads the speech model, then caches it on your machine. That is the only network request Lunosia ever makes.
Turn off your wifi. Pull the ethernet cable. Go offline completely.
Record, transcribe, summarise, export. Everything still works. A tool that phones home cannot do this, and no amount of policy wording changes that.
What it does
Built for the recordings people actually make.
Not a demo. These are the things that break in other tools, fixed because they broke here first.
Speakers across a room
A lecturer three rows away arrives far quieter than someone at the laptop. Most tools discard that audio as silence. Lunosia measures against the room's own noise floor instead of a fixed threshold, so distance stops mattering.
Voices over music
Speech with a backing track has no gaps between words, which fools silence detection into throwing the whole thing away. Lunosia hands it to the model anyway — and tells you when it's hearing music rather than showing you an empty page.
99 languages, or translated
Transcribe in the language spoken, or have any of the 99 come out as English text. French podcast in, English transcript out, still without a server.
A summary that quotes, not paraphrases
Key points, dates, deadlines, definitions and questions asked — pulled as real sentences that were really said. An exam date stays the exam date, because nothing rewrites it.
Check the text against the audio
Click any line to hear exactly that moment, at up to 2× speed. A transcript you can't verify is a guess you're asked to trust.
Fix a mis-heard name once
Models mishear technical terms the same way every time. Find and replace fixes all forty occurrences at once instead of one at a time.
Who it's for
People who can't send the recording anywhere.
Not a marketing segment — a hard requirement. For some of these, uploading audio to a third party isn't a preference, it's a breach.
Students
Record the lecture from the back row. Get the exam date, the chapters, and the deadline pulled out for you. Keep a whole semester without paying monthly.
Journalists
Interviews with sources who agreed to talk to you, not to a cloud provider's subprocessors.
Therapists and lawyers
Session and client notes where confidentiality is a professional obligation, and "we don't look at your data" isn't a good enough answer.
Anyone with an accent
99 languages, and translation into English when you need to understand rather than quote.
Pricing
One payment. Not a subscription.
There are no servers behind Lunosia, so there are no running costs to recover. Charging monthly would be rent, not a price.
Free
$0
- Record from your microphone
- All 99 languages
- Across-the-room mode for lectures
- Playback to check the text
- Saved transcripts, find and replace
- Export as .txt and .md
Pro
$19 once
- Transcribe files you already have
- Capture audio from any browser tab
- The Summary tab
- Subtitles as .srt and .vtt
- Speaker labels
- Save the recording as .wav
- The most accurate model
Questions
The things people ask first.
If there's no server, how does transcription work?
An open-source speech model (OpenAI's Whisper) runs directly in your browser using WebAssembly and your GPU. It downloads once — between 42 and 261 MB depending on which size you pick — and is cached afterwards. Your computer does the work that a cloud service would otherwise charge you monthly for.
What happens to my recordings?
Audio is held in memory so you can play a line back and check it, then discarded when you close the tab — unless you deliberately export it. Transcripts are saved in your own browser's storage. Nothing is uploaded, because there is nowhere to upload it to.
Is it actually accurate?
It's Whisper, the same model family many paid services run behind their APIs. The accurate model is genuinely good on technical vocabulary and names. It is not perfect — which is why every line is editable, clickable to replay, and searchable, so correcting it takes seconds rather than a re-listen.
Do I need an account?
No. There isn't one to make. Buying Pro gives you a licence key that unlocks the app on your machine — no email required, no login, nothing to be breached later.
How long can a recording be?
There's no limit on transcript length. Playback keeps the first 90 minutes of audio in memory, which covers a standard lecture or interview.
What if I lose my licence key?
Email your receipt and it gets reissued. Keys are tied to the purchase, so it's a quick lookup rather than a support ticket.
Try it before you decide anything.
The free version transcribes real recordings in 99 languages. No sign-up to get there.
Open Lunosia