Local-only dictation for macOS that writes like you meant it, and checks the model's work against what you actually said.
Hold a key, speak, release. Punctuated, cleaned-up text lands at your cursor, in whatever app you are using. Every model it uses runs on your Mac.
you say: the meeting is on tuesday at ten no wait wednesday at ten thirty in the big room
you get: The meeting is on Wednesday at ten thirty in the big room.
you say: so basically um the api returns a list of users and uh each one has an id
a name and an email and then you filter them by the active flag before
you render them
you get: So basically, the API returns a list of users, and each one has an ID, a
name, and an email. Then you filter them by the active flag before you
render them.
Tuesday at ten is gone, not punctuated into the sentence. The fillers dropped, "api" became "API", one breathless run-on became two sentences. Every example in this repository is real output, and you can reproduce any of them:
Kalamos --clean "the meeting is on tuesday at ten no wait wednesday at ten thirty"brew install --cask xmasyx/tap/kalamos
xattr -dr com.apple.quarantine /Applications/Kalamos.app # once: the app is not notarizedThe second line is needed once, because Kalamos is signed with the project's own certificate
and not notarized by Apple: without it macOS refuses to open the app (System Settings → Privacy
& Security → Open Anyway does the same thing). After that, Kalamos updates itself from its
menu: choose Check for updates… and it runs brew upgrade, clears the flag on the new copy
and relaunches. If you previously installed it with install.sh, switch to Homebrew with
brew install --cask --force xmasyx/tap/kalamos.
Alternatively:
curl -fsSL https://raw.githubusercontent.com/xmasyx/kalamos/main/Scripts/install.sh | bashRequires macOS 14+ on Apple Silicon (M1 or newer): transcription runs on the Neural Engine, which Intel Macs do not have. Builds are signed with the project's own certificate, not notarized: the installer clears the quarantine flag — read the script first, and see what macOS will say and why.
Removing it: install.sh --uninstall, or --purge to take the downloaded models
and settings with it.
It lives in the menu bar. What you reach for while working is one click down: the dictation language, the last transcription, the words you have taught it.
It checks its own output. The dangerous failure of a cleanup model is not a typo, it is the sentence that comes back reading better while meaning something else: a dropped condition, a missing "not", a rounded number. Those get refused rather than typed. No other dictation app does this, and it is the reason this one exists — the whole argument, with the failure cases.
Punctuation lands in 28 milliseconds, 118× faster than the language model it replaced, and it is the most accurate of every path measured. F1 against 40 sentences punctuated by hand from a style guide, with the time each path adds to a dictation:
| Punctuating with | Added time | . |
, |
? |
|---|---|---|---|---|
| Rules alone | 0 ms | 48.3 | 0.0 | 0.0 |
| A 7B language model | 3,300 ms 0.96 – 9.36 s |
72.6 | 53.8 | 50.0 |
| Whisper's own marks | 0 ms | 72.6 | 62.3 | 88.5 |
| Kalamos today | 28 ms never past 70 |
85.9 +18% |
78.4 +26% |
91.2 +3% |
The percentages are against Whisper's own marks, not against the language model, because the free option was never worse than the 7B on any mark and far better on commas and question marks. Improving on it costs 28 ms.
The second column is the one to read twice. The language model's cost grew with the sentence: on the bench corpus it ran from 0.96 s on a short one to 9.36 s on a long one, so the longer the thought you had just spoken, the longer you sat there. The classifier does not care — it stayed between 6 and 70 ms across the same corpus. A human starts noticing a delay at around 100 ms, so the punctuated text is simply there when you release the key, whether you said four words or eighty. Across twenty dictations a day, the old path spent 66 seconds waiting. This one spends 0.6.
And punctuation is only half of it. Whisper transcribes what you said, faithfully — which means it keeps every sort of, like, I mean, and so on. Measured over 386 real Italian dictations from one Mac's archive, comparing Whisper's raw output to what Kalamos typed: 180 were changed, 67 filler words were cut, and the delivered text came out 0.3% shorter while gaining all its punctuation. Spoken self-corrections are resolved by rule at 87.6% recall with zero legitimate words removed — the 7B model managed 70.1%, and to get there it deleted 64 words you had actually said and invented 44 you had not.
Long dictations keep their words. Speaking for two minutes without stopping used to cost you the odd word, in two different ways, both now fixed and both measured on a 120-recording archive: a last word said quietly was shaved off with the trailing silence (46 words lost, now 8), and a rare re-decode could drop a stretch in the middle (47 words out of 90 on one pass in eight, now zero in 24). Say the whole paragraph.
It also gave you 4 GB of memory back. Dictating used to hold Whisper and a 4 GB language model resident, about 5.5 GB of weights. Today dictation loads Whisper alone: 1.5 GB, measured as Activity Monitor measures it. That saving is the 7B not being there, not a change of speech engine. The language model is still there for the two jobs only it can do, rewriting a selection and translating as you dictate, and it is loaded when you ask for those and not before. How the 28 ms path is put together, and the two alternatives measured and thrown away.
It resolves what you take back mid-sentence. "We should ship on Friday, I mean Monday, because Friday is a public holiday" becomes "We should ship on Monday because Friday is a public holiday." The wrong day goes, the reason for it survives. Works the same in Italian and French.
It learns your words, two ways. A word misheard the same way every time is a correction, applied to the raw transcript before anything else. A term mangled differently every time is vocabulary, weighed in context by the model. Select a word anywhere on your Mac and press ⌃⌥L to teach it without opening a window: the dictionary grows while you work, and the transcription gets more precise as it does. A few other ⌃⌥ shortcuts do the same for the rest of the daily loop, and the menu bar lists them with their keys.
It rewrites text you already have. Select any text, hold the edit key, say how to change it, and it is rewritten in place. It translates on device: dictate in Italian and get English at the cursor, same model, nothing leaves the Mac.
Nothing is lost, and you can go back and listen. ⌃⌥V opens every dictation you have made, with search and filters, the audio next to the text, and playback at 0.5/1/1.25/1.5× with volume past the original when you spoke too quietly. Correcting a line there is not cosmetic: the correction feeds the words it learns from, so the next transcription is better for having been wrong once.
You can see it listening, and you decide where. A wave shows while the microphone is open: hanging from the notch as a band that continues the hardware, or as a small pill above the Dock. It is always draggable, from anywhere on it — drop it near either of those two spots and it snaps and stays there; drop it anywhere else and that is where it lives from then on. The wave follows the audio already being recorded; it listens to nothing extra.
And it fits the Mac you own. First run reads the chip, the memory, the cores and the free disk, then proposes the models that fit and shows you the figure that decided each choice. Accept and setup is two pages shorter.
There is no cloud path here. No account, no API key field, no cloud toggle, no server to point at. The single outbound request in the app's life is the model download on first run; after that it works with the network off.
Local-first is the polite way of saying a cloud path exists and you are not using it today. This is local-only, and you can prove it yourself with one grep: verifying the privacy claim. What touches your disk, precisely, is in what this project commits to.
Three speech engines, switchable in Preferences, all on device.
| Engine | What it runs | Pick it when |
|---|---|---|
| Whisper (WhisperKit, Core ML) — default | four model sizes from a menu | the default: it keeps the most words on a long dictation, and it learns your words before it guesses |
| Parakeet (FluidAudio) | one model, 461 MB | you want the smallest download and the fastest answer |
There was a third, a whisper.cpp engine, from 5 to 19 August 2026. It was added for two things Core ML could not do then, both were fixed somewhere else, and a head-to-head on twenty real dictations found nothing left to prefer. Which cleanup model your Mac can hold, and the measurements behind all of it, including what would bring that engine back: engines and models.
- Using it — the trigger, Edit Mode, spoken punctuation, translation, nothing is ever lost, troubleshooting
- Why it exists — the crowded shelf, the net under the model, and what is deliberately not here
- Engines and models — the three engines with numbers, and what your Mac can run
- Privacy — check the claim yourself
- What it commits to — enforced by code or checkable by you
- Architecture — how it is put together
git clone https://github.com/xmasyx/kalamos.git && cd kalamos
./Scripts/build-app.shNeeds full Xcode: the Metal shader compiler for MLX ships with it and not with the Command Line Tools.
Working, and used daily by its author. Not code-signed or notarized, hence the quarantine step. Three interface languages so far; adding one is mostly a matter of teaching it that language's Whisper hallucinations and self-correction markers.
Bug reports and pull requests are welcome. For a cleanup problem, attach the
--doctor output and what you said versus what you got.
Built on WhisperKit and FluidAudio for on-device speech, and MLX for the on-device language model. None of those projects are affiliated with this one.
MIT — see LICENSE.