NoType

Private, local dictation for macOS — speak naturally and keep typing.

⬇ Download for macOS · Releases · Features · Privacy · Build from source

NoType is a native macOS menu-bar dictation app. Focus a text field, start dictation, speak, and stop — NoType transcribes your voice locally and inserts the result back into the app you were using.

The downloadable build includes the multilingual Whisper large-v3 model and its tokenizer. There is no account, no remote transcription API, and no model download before your first dictation.

Quick start

1. Download NoType-0.3.0-arm64.zip and unzip it. The archive is about 1.4 GB because the local speech model is included.
2. Move NoType.app to /Applications. On first launch, Control-click the app and choose Open — the current build is ad-hoc signed and not yet Apple-notarized.
3. Allow Microphone access for recording and Accessibility access for global shortcuts and direct text insertion.

Then focus any text field and double-tap Command to start. Double-tap it again to stop, transcribe, and insert the result.

On the first launch, Core ML may need one or two minutes to specialize the bundled model for the Mac. Open NoType → Diagnostics to see Local Model: Preparing, Ready, or Failed without reading the debug log. Completed specialization is cached across normal app launches and Mac restarts.

If Control-click → Open is unavailable, clear the quarantine flag:

xattr -dr com.apple.quarantine /Applications/NoType.app
open /Applications/NoType.app

System requirements

• Apple Silicon Mac (arm64)
• macOS 15.0 or later
• About 4 GB of free space during download and extraction
• Microphone permission for recording
• Accessibility permission for global double-tap shortcuts and direct insertion

Why NoType

• Your voice stays on your Mac. Audio is transcribed through WhisperKit and Core ML; temporary recordings are deleted after each attempt.
• Chinese and English can share a sentence. Auto mixed recognition is designed for code-switching, with Chinese-first and English-first modes when you want a stronger bias.
• The model arrives ready. The release contains the large-v3 model and tokenizer, so first use does not depend on a separate setup or download.
• It returns to the right place. NoType remembers the focused input where dictation began. If that target changes, it keeps the transcript on the clipboard instead of inserting into the wrong field.
• It stays out of the way. No windows are required for normal use; status, modes, permissions, shortcuts, and diagnostics live in the menu bar.

Features

Dictation

• Start and stop from anywhere with Double Command, Double Option, or Command + Shift + H.
• Automatic five-minute recording limit prevents an abandoned session from running indefinitely.
• Very short accidental recordings are ignored instead of being sent through transcription.
• The transcription model is prepared in the background after launch to reduce first-use waiting.
• Diagnostics shows whether the local model is preparing, ready, or failed and offers a retry when preparation cannot complete.

Recognition

• Auto (中英混说) for natural mixed Chinese and English speech.
• 中文优先 and 英文优先 for language-biased recognition.
• Follow-model, Simplified Chinese, or Traditional Chinese output preferences.
• A separate shortcut cycles recognition modes without opening the menu.

Text delivery

• Insert and copy, insert only, or copy only after a successful transcription.
• Direct Accessibility insertion with a paste fallback for apps that do not expose a compatible text field.
• Captured-target protection prevents a completed transcript from landing in a different chat or document.
• Clipboard-preserving fallback when insert-only mode needs to simulate a paste.

Menu bar controls

• Live idle, recording, transcribing, inserting, and error states.
• Recognition and Chinese-script markers visible in the menu-bar icon.
• Configurable dictation and recognition-mode shortcuts that persist across launches.
• Permission status, the latest diagnostic event, and direct access to the local debug log.

Usage

1. Put the cursor where the transcript should appear.
2. Double-tap Command to start recording.
3. Speak normally. Mixed Chinese and English is supported in the default mode.
4. Double-tap Command again to stop.
5. NoType transcribes locally, returns to the captured app, and inserts or copies the transcript according to your selected success mode.

Open the menu-bar icon to change recognition mode, Chinese script, output behavior, shortcuts, or permissions. The default shortcuts are:

Action	Default	Other choices
Start / stop dictation	Double Command	Double Option, Command + Shift + H, Disabled
Cycle recognition mode	Command + Shift + Y	Command + Option + Y, Control + Option + Y, Disabled

Privacy

• Speech recognition runs locally through WhisperKit and Core ML.
• NoType does not use a remote transcription API.
• The release build does not download a model at runtime.
• Temporary audio is removed after transcription succeeds or fails.
• There are no accounts, analytics, or telemetry in the app.
• Microphone access is used only for active dictation.
• Accessibility access is used for the global modifier shortcut, focused-target capture, and text insertion.
• The clipboard is touched only when the selected output mode or an insertion fallback requires it.

Current release

NoType 0.3.0 is available as an Apple Silicon archive from the v0.3.0 release.

Artifact	Purpose
NoType-0.3.0-arm64.zip	Ready-to-run app with the Whisper model and tokenizer included
NoType-0.3.0-arm64.zip.sha256	SHA-256 checksum for download verification

The app is approximately 1.5 GB after extraction. The model is distributed as a GitHub Release asset and is intentionally not committed to this repository.

FAQ

macOS says NoType cannot be opened because the developer cannot be verified

The current release is ad-hoc signed and not Apple-notarized. Control-click NoType.app, choose Open, and confirm once. You can also run the xattr command shown in Quick start.

Why can the first launch take a minute or two?

The model is bundled, so NoType does not download anything. Core ML may still specialize the model for the current Mac the first time that model and compute configuration are used. NoType keeps running while this happens and shows the real state under Diagnostics → Local Model. Once Ready, the completed Core ML cache normally survives app launches and Mac restarts. An OS update, model change, app-bundle replacement, compute-configuration change, low-disk cleanup, or manual cache removal can require specialization again.

Why is the download so large?

NoType includes the Whisper large-v3 _turbo Core ML package and its tokenizer. The package includes an optional decoder-prefill model that reduces startup decoding work. This makes the archive much larger, but it also means transcription works locally on first launch without downloading a model or sending speech to a server.

Why does NoType need Microphone and Accessibility permission?

Microphone permission lets NoType record while dictation is active. Accessibility permission lets it observe the global modifier shortcut, remember the focused input, and insert the finished transcript. Copy-only output still needs Microphone access but does not require direct text insertion.

What happens if I switch chats or text fields while NoType is transcribing?

NoType tries to return to the input captured when dictation started. If that specific target is no longer available, it copies the transcript instead of risking insertion into a different conversation or document.

How do I uninstall NoType?

Quit NoType from the menu-bar icon, then move /Applications/NoType.app to the Trash. You can revoke its permissions under System Settings → Privacy & Security → Microphone / Accessibility.

Build from source

Requirements, build commands, and release packaging

Requirements:

• macOS 15.0 or later
• Apple Silicon for the current release-packaging path
• Swift 6.1 or a compatible Xcode toolchain
• A local WhisperKit Core ML model and tokenizer for a self-contained release

Run the test suite:

swift test

Build a small local app bundle:

./scripts/build_app.sh

This creates dist/NoType.app without copying the large model. It is useful for compile and bundle checks. A developer build without bundled resources uses the fallback paths configured in LocalWhisperPaths.swift.

Build the self-contained release archive:

WHISPER_MODEL_DIR="/path/to/openai_whisper-large-v3-v20240930_turbo" \
WHISPER_TOKENIZER_DIR="/path/to/whisper-large-v3" \
./scripts/build_release.sh

The packaging script builds the current checkout, bundles the model and tokenizer, signs the whole app bundle, and creates:

dist/NoType.app
dist/NoType-0.3.0-arm64.zip
dist/NoType-0.3.0-arm64.zip.sha256

The default signature is ad-hoc. To use an installed Developer ID identity:

CODESIGN_IDENTITY="Developer ID Application: Your Name (TEAMID)" \
./scripts/build_release.sh

The script verifies the signature with codesign --verify --deep --strict. Notarization is not automated yet; consult Apple's macOS distribution guidance before treating a build as a public, notarized release.

Project layout

• Sources/Typeless/Audio/ — microphone capture and temporary recorded clips.
• Sources/Typeless/Transcription/ — WhisperKit loading, language attempts, selection, and transcript cleanup.
• Sources/Typeless/Accessibility/ — focused-target capture, direct insertion, clipboard handling, and paste fallback.
• Sources/Typeless/Coordinator/ — dictation state transitions and orchestration.
• Sources/Typeless/App/ — menu-bar UI, permissions, diagnostics, and app lifecycle.
• Sources/Typeless/Hotkey/ — Carbon shortcuts and modifier double-tap detection.
• Tests/TypelessTests/ — unit coverage for transcription, coordination, shortcuts, state, permissions, and menu presentation.
• Packaging/Info.plist — app identity, version, permissions text, and minimum macOS version.
• scripts/ — local app and self-contained release builds.
• docs/decisions/ — product and engineering decision records.

Known limitations

• The downloadable release supports Apple Silicon only.
• macOS 15.0 or later is required.
• The current public build is ad-hoc signed and not notarized.
• There is no in-app updater; new versions are installed manually.
• The release model is fixed to Whisper large-v3; model switching is not exposed in the app.

Documentation

• Changelog — shipped user-facing changes.
• Known issues — understood limitations and future directions.
• ADR-001: Configurable shortcut input — why NoType uses curated shortcuts and modifier double-taps.
• ADR-002: Portable release packaging — why releases bundle the model instead of downloading it at runtime.
• ADR-003: Local model readiness — why Diagnostics reflects the actual Core ML loading lifecycle without a percentage.

Third-party terms

NoType uses WhisperKit. Its license is included at Vendor/WhisperKit-main/LICENSE.