Murmur
Voice dictation · Windows, Linux, Android

Hold a key. Speak. The words land where your cursor is.

Murmur turns what you say into clean, punctuated text and types it into whatever you are using: chat, email, a document, a code editor, a terminal. Fillers and false starts disappear, numbers come out the way you would type them, and nothing you said gets invented or lost.

v0.5.11 released 27 Sept 2026 · Free to start, no card. MIT licensed, source on GitHub.

What you said

um so the the invoice is one thousand two hundred dollars and it’s due on the fifteenth uh new line thanks abby

What landed where your cursor was

The invoice is $1,200 and it’s due on the 15th.

Thanks, Abby

Numbers
1200 · 15, same order
Words kept
11 of 12 content words
Command
“new line” applied
How it works

Three things, and none of them is a menu.

Murmur is a Wispr Flow-style dictation tool: one shortcut, your voice, and text that is ready to send.

  1. 01

    Hold a key, or tap the pill

    Ctrl + Win by default, any keys you like. Hold to talk and release to insert, or tap once for hands-free and tap again to stop. On Android the pill floats above the keyboard.

  2. 02

    Say it the way you would say it

    Fillers, stutters, false starts and “Tuesday, no, Wednesday” are yours to make. Spoken commands work too: new line, new paragraph, scratch that, quote and end quote.

  3. 03

    It lands where your cursor is

    Typed straight into the app you are in. Tone follows the destination: casual in chat, professional in email, exact identifiers in code editors, one line and no trailing period in a terminal.

The engine

The model does the language work. The verifier makes sure it did not cheat.

Most dictation tools either trust a language model completely or fence it in with rules that break on the next sentence. Murmur gives the model the raw transcript and then holds its answer to a few invariants that can be checked without understanding the text.

  1. 01

    Speech in, raw transcript out

    The speech model returns your words with their timings. When it stops early, which Whisper-style models do right after a word from your dictionary, Murmur transcribes the rest from the nearest pause and appends it, so nothing you said is dropped.

  2. 02

    The model sees the raw transcript

    Exactly as the speech model returned it, together with where the text is going, what is already before the cursor, your dictionary and your instructions. It removes fillers and false starts, applies spoken commands, writes numbers the way a person types them, lays enumerations out as lists and matches the destination.

  3. 03

    The verifier checks the answer

    Not empty, not chatty, not an answer to your question. Most of your words and roughly your length. Phrases that must stay verbatim are still there. And exactly the same numbers, in the same order, however the model chose to write them.

  4. 04

    A failed check never reaches your cursor

    One strict retry, then the rule-based cleanup of the transcript is inserted instead, and History tells you why. Your dictionary is re-applied after the model, and your snippets expand last.

Numbers are compared as values

The model may write a number any way a person would. It may not change, drop, merge or invent one.

SpokenTyped
five thirty pm5:30 pm
one million two hundred thousand dollars$1,200,000
version two point oh point one2.0.1
zero zero zero seven0007
five thousand, five thousand5,000, 5,000
Caught: for “five thousand, five thousand” the model wrote a single 5,000. Signature 5000 5000 became 5000, the answer was refused, and the strict retry kept both.
Your models, or Murmur's

Bring a key, run a local model, or sign in and skip all of that.

Every build works on its own with a speech model you connect. An account adds Murmur's models and keeps your settings the same on every device. Switch between the two per device, any time.

Your own provider

Local mode. Free, forever, no account.

  • Any OpenAI-compatible speech endpoint: OpenAI, Groq, Mistral Voxtral, a local whisper server
  • Deepgram and ElevenLabs natively, with your dictionary sent as key terms
  • One-click model discovery, a fallback model and a built-in latency test
  • Keys stay on the device. No account, nothing is uploaded.

Murmur's models

With an account, while the server is open to it.

  • Speech and formatting models included, nothing to configure
  • Dictionary, snippets, style and stats in step on desktop and Android
  • Dictation history sync, opt-in, because it contains what you said
  • Works offline from a local mirror; edits queue and replay in order
Platforms

A desktop app for Windows and Linux. A native app for Android.

Both check GitHub Releases for updates, verify every download against the release’s checksums and install it when you are not dictating. You download Murmur by hand exactly once.

Windows

Windows 10 and 11, x64 and arm64. Hold the shortcut in any app; the tray shows what Murmur is doing.

Installer for both architectures, per-arch installers, a portable .exe

Download for Windowsv0.5.11 · 217 MB

Linux

GNOME, KDE and friends on X11 or Wayland, x64 and arm64. Follows your desktop’s accent colour.

AppImage and .deb, both architectures

Download for Linuxv0.5.11 · 145 MB

Android

Android 8.0 and up. A floating pill above the keyboard dictates into any app, Material You throughout.

APK, installs its own updates

Download for Androidv0.5.11 · 37 MB

macOS builds (Apple silicon and Intel) ship with every release as well, marked experimental. Every file, with checksums.

On your terms

Dictation hears everything you say. Murmur is built like it knows that.

Four commitments about where your words go, each one checkable in the source rather than promised in a policy.

Recordings never leave the device
Every dictation’s audio stays with its History entry on your machine so you can play it back or retry a failed one. Nothing is uploaded except the clip being transcribed.
Keys and model choices are device settings
Which speech model a device uses and any API keys of your own are never synced. An account syncs your dictionary, snippets, style and stats, and history only if you ask.
Updates you can check
Both apps update from GitHub Releases and verify each download against the release’s SHA256SUMS.txt. Releases without one are shown but never installed unattended.
Open source, MIT
The desktop app, the Android app, the text engine and the backend are one repository. Read the prompt the model gets and the checks its answer must pass.

Read the source

Start dictating in about a minute.

Download, hold the shortcut, say something. Connect a model of your own or sign in for Murmur’s.