Turn your voice into finished text. Works in Slack, Gmail, Word, your terminal and any other app.
Speech in, text out
Nobody talks in capitals and full stops. Murmur hears the run-on and types the sentence.
Free · Windows 10 1809 or later, 64-bit Intel or AMD · Nothing leaves the machine
There is no integration to install and no plugin to approve.
If the window takes a cursor, it takes your voice.
Everywhere
Murmur pastes into whatever had focus when you pressed the key. Pick a destination and watch the same four seconds of speech land in it.
What's inside
The model sits on your disk and runs on your processor or your graphics card. No Wi-Fi, no problem.
From instant-but-rough to slow-and-precise, downloaded on demand and kept afterwards. All of them free, forever.
Commas, full stops and capitals arrive without being dictated, because the decoder is a language model.Say it, don't spell it.
Turkish included. Detection is automatic, but naming your language in the menu is 2.6x faster and cannot guess wrong. It will translate them all to English too.
Reminder to send the quarterly numbers to Dilek before Friday.
➤No copying, no switching window. The transcript goes straight into the field the cursor was already in.Your hands never leave the work.
4.7× realtime on an RTX 5070 Ti with the recommended model; 1.5× on the processor alone. The text is waiting before you are.
The History tab keeps the session's transcripts so you can grab one again. Gone when you quit, unless you ask it to keep them.
Point it at a recording with --file and get the text, no microphone needed.
Press once to start, once again to stop and paste. Esc throws the take away.
Ctrl+Space to begin with, changed in the setup guide and kept in config.json.
A voice-activity filter runs first, so an empty room never produces an invented sentence.
Integrations
Slack, Cursor, Notion, Word, your terminal.
You say it, Murmur puts it there.
Set it up once
Five screens on first run, then a window you rarely need to open again.
Six models between instant and immaculate. Switch whenever the job changes; the download is kept.
The installer ships the processor build. One button on the Speed page adds the CUDA libraries, and it needs no administrator rights.
Ctrl+Space is only the default. Per-application paste overrides are there for consoles that want Ctrl+Shift+V.
Beep boop
Use Murmur with Claude Code, Cursor, Codex or any other agent that takes a prompt, without touching your keyboard.
Any editor or terminal that accepts a paste, which is all of them.
People speak roughly three times faster than they type. Give the model the whole picture instead of the short version.
Models
Bigger is more accurate, smaller is faster. They download on demand and are kept afterwards.
| Model | Size | Languages | Best for |
|---|---|---|---|
| Tiny | 75 MB | 99 | Short commands and clear speech in a quiet room. |
| Base | 145 MB | 99 | A sensible floor on machines with no graphics card. |
| SmallRecommended | 480 MB | 99 | Everyday dictation. Text comes back before you've thought of the next sentence. |
| Distil Large v3 | 1.5 GB | English | The best quality-per-second here, if you only dictate in English. |
| Medium | 1.5 GB | 99 | The multilingual step up. Better with names, jargon and noisy rooms. |
| Large v3 | 3.1 GB | 99 | Strong accents and difficult audio. The slowest of the set. |
Get it
Open source, no subscription, no account. The speech models are MIT licensed and cost nothing to run.
Windows will say it protected your PC - the installer is not signed. Choose More info → Run anyway.
Anywhere you can paste. Two exceptions worth knowing: windows running as Administrator will not accept the paste, because Windows blocks input across privilege levels, and a few older consoles want Ctrl+Shift+V instead of Ctrl+V, which you can set per application.
On a recent NVIDIA laptop GPU the recommended model transcribes four to five times faster than you speak, so a four second sentence returns in under a second. Without a graphics card, expect roughly real time on the smaller models.
No. Every model is MIT licensed and publicly hosted. You download once and run locally forever, with no per-minute charge and no account.
No, and it will say that. The installer is not code-signed, so SmartScreen warns about it and puts the run button behind More info → Run anyway. The warning means the file has no certificate and few people have downloaded it yet, not that anything was found in it, a certificate costs a few hundred pounds a year, which a free tool does not have. If you would rather not take that on trust, the whole program is on GitHub and you can run it from source instead.
Yes. Every model except Distil Large v3 covers 99 languages and will detect yours automatically, but name it if you can. Detection reads the opening seconds and commits, and a wrong guess does not produce a translation, it produces confident nonsense, because the decoder picks the wrong vocabulary and then writes fluently in it. Naming the language also skips the detection pass: on the same eight-second English clip here, 0.73s named against 1.88s detected, 2.6x faster. It is in the tray and pill menus as well as on the Language page, so switching between two languages is a right-click, not a trip into settings.
No. Audio exists in memory only for the seconds between pressing the shortcut and the text appearing, then it is discarded. There is no code path in Murmur that writes a recording to a file.
Four files, all in %LOCALAPPDATA%\Murmur and none of them sent anywhere. config.json is your settings. stats.json is totals, words, seconds, and which applications you dictated into; no transcript and no text of any kind reaches it. murmur.log is startup and error output, kept because the installed build is a windowed program with no console to print to; it holds no transcripts, but the file paths in it contain your Windows username, so read it before pasting it into a bug report. history.json only exists if you tick keep these after Murmur closes on the History page, which is off by default; it holds the last 300 transcripts as plain text, and the Clear button on that page empties it. Leave the tick off and your transcripts are gone the moment Murmur quits.