Vocs

Press a key, talk, press it again — your words are on the clipboard and pasted before you can look up.

Local-first speech to text for macOS. No account, no subscription, no word cap, and no audio ever leaves your machine.

Download · Releases · Changelog · ♥ Sponsor

---

What it does

• Global hotkey — ⌥Space anywhere. Toggle, or hold-to-talk.
• Transcribes while you speak, not after you stop, so finishing is as fast after a five-minute monologue as after a five-second one.
• Pastes into whatever had focus, and puts your clipboard back afterwards.
• Escape to cancel, with a three-second countdown. Press Escape again and the recording resumes with nothing lost.
• 99 languages, auto-detected or pinned.
• A custom dictionary for the names and jargon the model gets wrong — used both to bias recognition and to correct the output.
• A notebook of everything you dictated — Dictations — searchable and filterable by period and status. Open one to edit it like a notepad: find in it, rename it, and it saves itself. It is local history, so a failed paste is never a lost thought. Where the rules changed your words, History shows what you said against what was delivered, and Use original takes your wording instead. Copy Last Transcript in the menu bar, or an optional global shortcut (off until you turn it on), is the way back from a paste that went nowhere.
• Writing style — natural, casual, formal or literal, set once for everywhere.
• Preview before pasting and live captions (both off by default): see the finished text for a moment before it lands, or watch what Vocs hears rise out of the recording pill as you speak.
• Take words back by voice. Say "scratch that", "undo last sentence" or "scratch everything" and the words are removed from what gets pasted (English).
• A language per hotkey. Give each language you speak its own shortcut and skip auto-detection entirely. Off until you turn it on.
• A dictation scratchpad. Ramble into a floating notepad, press Return to turn it into bullets, a task list or a clean paragraph, and ⌘Return to send it to the app you came from.
• Filler removal that does not eat real words, at three strengths. "I like pizza" is never touched; "it was, like, fine" loses the "like".
• Optional grammar fixes for slips that are wrong every time — "a idea", "she don't", "we was", "could of been" — with the rest of your sentence left exactly as you said it. Off by default; quoted speech and your dictionary are never touched. That rule set is small on purpose: it fixes less, and it cannot make a mistake you would not have spotted.
• Smart grammar, on your Mac. Turn on an optional on-device model (a ~120 MB download, then offline) for the agreement and tense slips rules cannot see. It never rewrites you: only safe word-level fixes are kept, and names, numbers, code and your dictionary are off limits. It also learns: a fix you keep three times and never undo starts applying itself, and you can see, switch off or delete everything it has learned.
• A context line for the model — jargon, product names, the topic — that it listens for ahead of your dictionary.
• A live microphone meter in Settings → Recording, on the device dictation will actually use.
• Back up and restore your dictionary as one JSON file, to move them between Macs. Importing adds; it never deletes.
• Crash-safe: the transcript is saved as it grows, so a crash keeps what you had said.
• Honest stats, including the p50/p95 latency the app actually achieves, per stage: capture, recognition, correction, output.
• Export diagnostics — a plain-text support file you create and send yourself. There is no telemetry.

Everything runs on your Mac. The only network requests Vocs ever makes are the first-run model download and an optional update check, and both are visible and both can be turned off.

Requirements

• macOS 13 or later, Apple Silicon or Intel
• ~600 MB of disk for the default model

Running it

pnpm install
pnpm tauri dev

First launch opens setup: grant the microphone, download the model, and test the hotkey. After that Vocs lives in the menu bar. Opening the app (from Finder, Launchpad or Spotlight) shows the dashboard; when it starts by itself at login it stays quietly in the menu bar.

Building a release

pnpm tauri build

The bundle lands in src-tauri/target/release/bundle/ — a .app, a .dmg, and an updater archive.

Signing the updater archive. Because a public key is configured, Tauri also signs the update artifact, and it needs the matching private key:

TAURI_SIGNING_PRIVATE_KEY="$(cat ~/.vocs-updater.key)" \
TAURI_SIGNING_PRIVATE_KEY_PASSWORD="" \
  pnpm tauri build

Both variables are required. Setting the key without the password makes Tauri prompt for one interactively, which fails in any non-interactive shell with a misleading "incorrect password" error — the password is empty, it simply had no way to tell you so.

Without either, the .app and .dmg still build correctly and only the update artifact goes unsigned. In CI both come from repository secrets.

Back that private key up somewhere you will still have in five years. It is the only thing that can sign an update. If it is lost, every existing installation is permanently stranded on whatever version it has — there is no recovery, because the public key is compiled into the apps already out there.

Distributing without an Apple Developer account

Vocs is distributed as a DMG from our own website, not the App Store, and without Apple notarisation. That works, with two consequences you must plan for.

What a user sees on the first launch. Gatekeeper blocks the DMG itself ("Vocs-macOS.dmg" Not Opened: "Apple could not verify…"), and may block the app again once it is copied out. The user chooses Done (not Move to Trash), then opens System Settings → Privacy & Security, scrolls to the message and chooses Open Anyway (or runs xattr -dr com.apple.quarantine on the DMG, and on /Applications/Vocs.app if asked). This happens once per file. In-app updates are downloaded by the app itself, so they do not trigger it again. The download page on the website must show these steps.

Sign every release with the same self-signed certificate, forever. An ad-hoc signature is new on every build, so each update would silently reset the user's Accessibility grant and pasting would stop working. A stable certificate keeps the grant. Create it once:

1. Keychain Access → Certificate Assistant → Create a Certificate… — name it Vocs Signing, identity type Self Signed Root, type Code Signing.
2. Export it with its private key as a .p12 and keep it in a password manager. Losing or replacing it voids every user's Accessibility grant.
3. Add three repository secrets: APPLE_CERTIFICATE (base64 -i vocs.p12), APPLE_CERTIFICATE_PASSWORD, and APPLE_SIGNING_IDENTITY (Vocs Signing). Leave APPLE_ID, APPLE_PASSWORD and APPLE_TEAM_ID unset; the release workflow then signs without notarising.

The updater needs its own, separate keypair (pnpm tauri signer generate); its private half goes in TAURI_SIGNING_PRIVATE_KEY.

If an Apple Developer account is added later, the only change is the secrets: a Developer ID certificate in APPLE_CERTIFICATE and the three Apple variables. Existing users then need to re-grant Accessibility once, because the signing identity changes.

Permissions, and what happens without them

Permission	Needed for	Without it
Microphone	Recording	Vocs cannot record. Required.
Accessibility	Pasting for you	Everything still works — text goes to the clipboard instead.

macOS shows each of these prompts once per app. If you dismiss one, the dialog never comes back, so Vocs takes you straight to the right System Settings pane instead of asking again.

If Accessibility keeps forgetting you granted it

macOS keys permission grants to the bundle id plus the code signature. Builds here are ad-hoc signed, which produces a new signature every time, so every rebuild looks like a different app and the Accessibility grant is lost. Pasting silently degrades to clipboard-only and nothing explains why.

That is a development annoyance, not a bug, and it does not affect a released build. If you are rebuilding often, create one self-signed code-signing certificate in Keychain Access and sign every build with that same identity — the signature then stays stable and macOS stops forgetting. docs/03 §3.4 has the detail.

For developers

The architecture is documented in docs/, and CLAUDE.md is the rulebook. Start with docs/00-START-HERE.md; read docs/06-CONVENTIONS-AND-GREP.md before writing your first file, because navigation here works through a comment-based search index rather than by reading the tree:

pnpm sot SessionState      # which files own this symbol
pnpm sot:show AudioChunk   # the same, with each file's header

Two ideas carry the codebase. The registry (src-tauri/src/registry/) is one table describing every feature — adding an entry wires up its settings UI, permission preflight, nav placement, hotkey and metrics at once. The command factory (src-tauri/src/ipc/factory.rs) is the single function every IPC call passes through, so validation, permissions, reentrancy, tracing and error mapping have one implementation instead of one per handler.

Types cross to TypeScript through tauri-specta: src/lib/bindings.ts is generated by cargo test and must never be hand-edited.

cargo test --manifest-path src-tauri/Cargo.toml   # also regenerates bindings
pnpm typecheck

Contributing

Issues and pull requests are welcome. Read CONTRIBUTING.md first, and report security problems privately as described in SECURITY.md. Everyone taking part follows the Code of Conduct.

Licence

MIT. See LICENSE. Third-party licences, including the models Vocs downloads, are listed in THIRD_PARTY_NOTICES.md.