Talking is the fast part. Typing is the second pass — and so is cleaning up what a transcriber hands back. Hold a key, say it, and finished text lands in whatever field the cursor is already in: a chat box, an email, a commit message, a prompt.
On your phone? Ask for a seat, or see what the tidying does →
This demo is drawn in code, not recorded — the capsule here is the one you get in the app.
A few dozen inputs a day, and a minute or two saved on each one. That is the whole argument this page is making.
Say it, get it written. The “um”, the “you know”, the false start get cleaned out on the way; what lands is a sentence you would have typed.
Right ⌘Speak English, Chinese lands; speak Chinese, English lands. The direction is worked out for you, so a message to a colleague in another language is one held key.
Right ⌥Say the question and what lands is the answer, not your question. A flag you half-remember, a conversion, a spelling — without leaving the window.
fnSelect a paragraph and say how it should change — what lands back is the changed version. Being able to revise by mouth is what closes the loop.
Right ⌃You are not writing the code character by character any more — you are explaining what you want. “Make this function async, then put a retry around it.” Three seconds to say; thirty to type. Murmur has no plugin, no integration and nothing to set up — it is system-level dictation, and it lands at any cursor, including the prompt in your terminal that is waiting for input.
We put the recognition models we could get our hands on through one internal evaluation and kept three. None of them wins everywhere — how you talk is what decides it — so they are split by situation: strengths and weaknesses printed on the front, one menu to switch, no terminology to learn first.
Standard is the tier open to everyone today; the other two are opening up through the closed beta. “2× quota” means exactly that: on that tier a minute of long recording takes two minutes out of the monthly pool, and dictation draws from your weekly allowance at twice the rate. The multiplier is printed next to the tier; there is no hidden conversion. Lose the network or run out of allowance and it falls back to on-device recognition by itself: slower, weaker when you mix languages, but never a brick, and never a lost sentence.
Accounts, sign-in, subscriptions — that is the entire job our server does. Your recordings, transcripts, history and dictionary are never uploaded; they exist on your Mac only. For cloud recognition the audio goes straight from your Mac to the provider you picked, not through us. Choose on-device recognition and not one byte leaves this computer.
You speak — the audio leaves from your Mac and nowhere else
~/Library/Application Support/Murmur/
The recordings, and every word you have said, are in that folder. Open it whenever you like. Delete it and it is actually deleted.
Every dictation keeps the audio and a history entry. Cloud down, recognition wrong? Open the history and run it again on another engine. No path through this app makes a sentence you said disappear.
If the translation did not happen, it tells you it did not happen. If the microphone heard nothing, it says so on the spot. It will not quietly push some fallback mess into the window you are typing in.
On-device: not one byte leaves this computer. Cloud: which provider, and what got sent, is written in Settings in plain words — not on page eight of the terms.
The “release the key to text on screen” latency is a measured median, not a number a copywriter picked. Each engine's weak spots are printed next to its strong ones — switch if it does not suit you; your ears decide.
Long recording is a separate pipeline. The caption window floats above any full-screen window — the talk keeps playing, the captions keep up; turn on the second column if you want a translation beside them. Captions are an expendable preview: if they drop, the recording does not. The full transcript lands in your history when you stop, with the speakers told apart.
Drawn in code, not recorded. Free tier: 30 minutes a month. Pro: no limit on length.
Your recordings, transcripts, history and dictionary all live on your Mac; our server only handles accounts and subscriptions. For cloud recognition the audio connects straight to that one recognition service and does not pass through our server — that holds while you are on the free allowance we pay for, too. One exception, stated plainly: on the free allowance the text for the cleanup step goes through our server before it reaches the model provider. Pick on-device recognition and nothing leaves the computer at all. Proper nouns from your dictionary are sent along with the request, so that they get heard correctly — all of this is written in Settings, not buried in terms.
Offline, or out of allowance, it falls back to on-device recognition by itself — slower, weaker when you mix languages, but your words are not lost. Once you are back, any entry in the history can be run again on a cloud tier.
English, Chinese, and the two of them mixed inside one sentence — that last case is the one it was built to take seriously. Chinese with a regional accent (Cantonese, Sichuanese, Wu) has a tier of its own, in closed beta and rolling out.
No. Cloud allowance is invite-only for now: get a seat, sign in, and the free allowance is there — it is a cloud bill we pay for you, and when it runs out the app returns to on-device recognition rather than stopping to ask you for money.
On your phone? Ask for a seat, or see what the tidying does →
Requires macOS 26 or later on an Apple Silicon (M-series) Mac · Free to start · Cloud allowance by invite
On an Intel Mac, or still on macOS 25 or earlier? It won't install yet — leave your email at foraiandfocus+murmur@gmail.com and we'll tell you the moment it's supported.
The download is free and on-device recognition works without an account. The cloud allowance is invite-only for now — email foraiandfocus+murmur@gmail.com and tell us in one line what you'd use it for.
Unzip it and drag Murmur into Applications. On first run it asks for the microphone and for accessibility — the first to hear you, the second to put the words at your cursor.