Tap fn to dictate where your cursor is.
macOS 14+ · Apple Silicon only · English only · initial 1.26 GB model download
Tap Fn once and speak. Tap again and the text lands in the app you were already using — recorded, recognized, and cleaned up entirely on your Mac. Ordinary text fields get the paste; a terminal, a code editor, a password field, or a target Voiceour could not read gets the transcript on the clipboard instead, and no setting widens that. Permissions and delivery safety has the matrix. The mercury recording island is synchronized to the display it occupies: 120 fps on a ProMotion screen, the native rate on lower-refresh displays, capped at 120.
Terms you teach in the Glossary do more than fix spelling: Voiceour also repairs close phonetic mishearings of them, by deterministic matching against your own terms rather than a model rewriting your words. Terms that are ordinary English words are held out of that phonetic step, so a word you actually said is never traded for one that merely sounds like something you taught.
No account, no telemetry, no cloud transcription. The only network request fetches the recognition model, pinned to ggml-org/parakeet-GGUF revision 35156454d1a39de06863303dd209fd2bed6ee079. Settings offers a Compact version of that same model — a 0.67 GB download and about 1.34 GB cached after first load, instead of Balanced’s 1.26 GB download and about 2.51 GB cache. Their accuracy matches to within what the benchmarks can resolve, and which of them decodes faster has changed with the macOS version, so treat it as a footprint choice; it applies the next time Voiceour starts.
Home keeps the running totals: how long you have spoken, how much typing it saved, and which apps the words went to.
Voiceour is MIT licensed, so you're free to read it, fork it, and build your own version.
It's maintained by one person, so I keep the scope tight and won't merge every pull request. Please open an issue before starting anything large, so we can agree on the approach first. Bug reports and ideas are always welcome.
Two things that are especially welcome:
- Documentation fixes. Wrong commands, dead links, anything out of date.
- Fixes under
Vendor/parakeet/. That directory is a copy of ggml-org/whisper.cpp. Send those upstream instead — they'll reach every project using it, and they'll survive the next time Voiceour updates its copy.
CONTRIBUTING.md has the checks to run before opening a PR.
There is no signed release yet, so build from source. You need macOS 14 or newer on Apple Silicon and a Swift 6 toolchain — Xcode 16 or newer, or just its Command Line Tools.
git clone https://github.com/joswha/voiceour.git
cd voiceour
make buildmake on its own prints every target.
Start on the fake backend. It needs no model download and no permission grant: synthetic audio through the real pipeline, real UI.
make devThen run the real app. This assembles .build/Voiceour.app, downloads the pinned model on first launch — 1.26 GB, with progress in the menu bar — and asks for Microphone at the first recording, plus Accessibility if you want the paste rather than a clipboard copy. The first launch opens the console on Home, where a first-run card states the tap gesture, the download's progress, and which of those two permissions is required — it retires itself once you have dictated once. After verification, first load also builds a model-sized acceleration file locally; it is not a second download.
make runThat bundle is ad-hoc signed, not notarized, so its code identity can change when you rebuild and macOS may ask for both grants again. Run make signing once for a stable local identity; developer setup covers signing, notarization, and the rest of the commands.
Why Apple Silicon only?
The vendored parakeet.cpp/ggml code is arm64-only, and Vendor/parakeet/ggml/embed/ggml-metal-embed.c enforces that at compile time. Supporting Intel would mean vendoring upstream's x86 sources and splitting Package.swift into per-architecture targets, since SwiftPM has no architecture build condition.
It's a deliberate choice, not something on the roadmap — see non-goals.
What network requests does Voiceour make?
Exactly one: downloading the recognition model. Transcription itself is entirely local. No telemetry, no account, no analytics, no update checks.
That request goes to huggingface.co, for the repository ggml-org/parakeet-GGUF at revision 35156454d1a39de06863303dd209fd2bed6ee079, and fetches one of two files:
| file | download | when |
|---|---|---|
ggml-parakeet-tdt-0.6b-v3-f16.bin |
1.26 GB | Balanced (default) |
ggml-parakeet-tdt-0.6b-v3-q8_0.bin |
0.67 GB | Compact |
On first load Voiceour creates a second, model-sized local acceleration file from the verified weights. It is generated locally, never downloaded. Total cache use is about 2.51 GB for Balanced or 1.34 GB for Compact; switching variants removes the other variant’s cache directory.
The revision is pinned so you can fetch and hash those exact bytes yourself, and Voiceour verifies the download against a SHA-256 built into the app before loading it.
Worth being clear about one thing: like any download, that HTTPS request tells Hugging Face and its CDN your IP address and User-Agent. Local transcription doesn't change that.
You can block Voiceour in a firewall such as Little Snitch if you'd rather it didn't happen. Without the model there's no transcription, but the fake backend still works, since it downloads nothing.
Where does Voiceour keep my data?
On your Mac, and nowhere else. Audio is never saved — the recording is deleted as soon as it's transcribed. Transcripts stay in a local history file capped at the newest 500, and a secure field like a password box records nothing at all.
Settings has a Clear History button that erases the transcripts and the lifetime counters together. Permissions and delivery safety covers the rest.
Which languages does it support?
English only. The model is parakeet-tdt-0.6b-v3 and there's no language picker — see non-goals.
Is there a Homebrew install?
Not yet. A Homebrew cask has to pass macOS Gatekeeper checks, which needs a notarized build, and notarizing needs a paid Apple Developer account this project doesn't have. Building from source is the supported path for now.
The build fails on an older toolchain
Swift 6 is required: Xcode 16 or newer, or just its Command Line Tools.
macOS keeps asking for Microphone or Accessibility after a rebuild
macOS ties a permission to the app's code signature, and an ad-hoc development build gets a new one each time. Run scripts/setup_local_signing.sh once for a stable local certificate; scripts/bundle.sh picks it up automatically.
A permission is stuck
Quit Voiceour, reset it, and it'll ask again on the next recording.
tccutil reset Microphone com.voiceour.app
tccutil reset Accessibility com.voiceour.appDictation pastes into some apps but only copies in others
That's deliberate. Ordinary text fields get the paste; terminals, code editors, password fields, and any target Voiceour couldn't inspect get the transcript on the clipboard instead, and no setting widens it. Permissions and delivery safety has the full matrix.
The three-bet Apple-Silicon study is published with its evidence and rejected paths in the final research report:
- Technical vocabulary: deterministic repair and the research-only multi-model teacher raised exact jargon recall from 161/346 to 295/346 while holding the general transcript set unchanged. The ensemble is an evaluation and distillation tool, not a shipping configuration: its artifacts exceed the product's storage contract.
- Neural Engine execution: the 6/8/15-second CoreML encoder ladder cut measured compute energy by 57.8%; contamination-resistant ABBA measurement showed a 43.2% reduction. Promotion still needs another Apple Silicon SoC, hosted artifacts, and cold-specialization validation.
- Quantization: current-toolchain q8 was quality-non-inferior, tied or faster, and reduced the model download by 587 MB. CPU-tail-only q8 halved the tail weights; broader mixed quantization and margin-only quality prediction did not clear their gates.
These are M4 Pro engineering results, not unqualified product claims. The research index links the preregistrations, negative results, measurements, and remaining promotion gates.
Architecture · Permissions and delivery safety · Benchmarks · Research · Non-goals · UI design · Developer setup · Contributing · Changelog
MIT licensed. That covers Voiceour's own code; third-party components keep their own terms. NOTICE credits the vendored sources and the speech model, and benchmarks/DATA-LICENSE.md covers the CC BY 4.0 corpus transcripts quoted in the committed benchmark reports.
Voiceour is an independent macOS dictation app, not affiliated with, sponsored by, or endorsed by Apple Inc.