Worried about voice privacy?
Speech recognition and AI cleanup run on your Mac instead of sending daily dictation to a cloud transcription service.
PRIVATE · LOCAL · AI
Private AI voice typing for Mac. Your voice stays local.
Hold a shortcut to speak and release it to type in any app. CoveType runs Qwen speech-to-text and AI polishing locally, detects 30 languages, and supports Apple on-device translation.
- macOS 15+
- Apple silicon (M1+)
- No account
- Free & open source
Listening
Release to transcribe and type
Processed locally
BUILT AROUND REAL PROBLEMS
AI dictation without the usual privacy and workflow trade-offs.
Cloud voice typing can expose sensitive speech, add network delay, produce messy text, and interrupt everyday shortcuts. CoveType is designed around those pain points.
Waiting on cloud results?
Local models remove the round trip to a remote ASR or writing API and continue working without an internet connection.
Raw transcripts need cleanup?
An optional local Qwen model improves punctuation, sentence breaks, and spoken phrasing before text reaches your app.
Shortcuts trigger by accident?
Record your preferred key or key combination and set a hold delay so shortcuts such as Control-C keep behaving normally.
CoveType compared with typical cloud AI voice typing
| Capability | CoveType | Typical cloud dictation |
|---|---|---|
| Speech-to-text | Runs locally on your Mac | Audio is uploaded for processing |
| AI text polishing | Local Qwen model | Often handled by a cloud LLM |
| Translation | Apple on-device framework | Often a remote translation API |
| Account | Not required | Commonly required |
| Source code | Open source under GPLv3 | Usually proprietary |
WHAT IT DOES
Voice input without the cloud detour.
Local transcription
Qwen3-ASR 0.6B 8-bit runs on Apple silicon and automatically detects 30 languages.
Local AI polishing
Qwen3.5 0.8B 4-bit can improve punctuation, sentence breaks, and spoken phrasing without a cloud writing API.
Instant translation
Speak one language and output another through Apple Translation and downloaded on-device language packs.
Custom hold-to-talk
Record a key or key combination and set the hold duration so normal shortcuts do not trigger recording.
Works across apps
Release the shortcut and CoveType inserts the result at the cursor in your current application.
Made for the menu bar
A small breathing lamp shows idle, listening, processing, and completion states. CoveType can start at login.
PRIVACY & SECURITY
A local processing path you can understand.
Daily transcription and AI polishing do not call cloud ASR or cloud writing APIs. Recording starts only after your configured shortcut is intentionally held.
Audio stays local
Temporary recordings are processed on the Mac and cleaned up after the workflow finishes.
Models stay local
Qwen3-ASR and Qwen3.5 are stored in CoveType’s local application data.
No account or sync
CoveType does not need a login and does not synchronize your transcript history.
Permissions are limited
Microphone captures speech; Accessibility listens for the global shortcut and inserts text into the active app.
When is a network connection used?
The first installation downloads the pinned runtime and models. macOS downloads Apple Translation language packs when needed. A configured update check may contact the release endpoint. Normal transcription and polishing remain local.
Your local voice-to-text path
- 1SpeakRecording starts only while your chosen shortcut is held.
- 2Transcribe locallyQwen3-ASR converts speech to text on the Mac.
- 3Polish locallyOptional Qwen3.5 cleanup improves readability on device.
- 4Translate or typeUse Apple on-device translation or insert the result into the active app.
OPEN SOURCE BY DESIGN
Privacy claims you can inspect—not just trust.
CoveType’s macOS application code is public under GPLv3. Review how recording starts, where models run, when networking is used, and how text is inserted into other apps.
View source code on GitHubAuditable
Inspect the Swift client, local AI worker, installer, update channel, and permission handling.
No account
No sign-up, subscription identity, or transcript sync is required for everyday voice typing.
Clear network boundary
The first install, translation language packs, updates, GitHub feedback, and source links can use the network. Daily transcription and polishing stay local.
LANGUAGE SUPPORT
30 languages, detected automatically.
This is the exact recognition list reported by the installed Qwen3-ASR model. Cantonese is listed separately from Chinese. You do not need to select the spoken language before recording.
ChineseEnglishCantoneseArabicGermanFrenchSpanishPortugueseIndonesianItalianKoreanRussianThaiVietnameseJapaneseTurkishHindiMalayDutchSwedishDanishFinnishPolishCzechFilipinoPersianGreekRomanianHungarianMacedonian
Translation output is a separate capability.
The recognition list above does not guarantee the same 30 translation targets. Available Apple on-device translation pairs depend on the macOS version, region, and language packs installed by the user.
DOWNLOAD & INSTALL
One guided installer for Apple silicon Macs.
Requirements
- Computer
- Mac with M1 or newer Apple silicon
- System
- macOS 15 or later
- Memory
- 8 GB workable; 16 GB recommended
- Storage
- At least 5 GB free
- First install
- Internet connection and an administrator password
- 1Download and unzip
Open the CoveType installer ZIP and extract it.
- 2Run the installer
Open “Install CoveType.command”. It installs the app, isolated runtime, models, and login item.
- 3Approve two permissions
Follow the system-language guide to enable Microphone and Accessibility. The installer checks both.
CoveType 2.1.4macOS 15+ · Apple silicon · Installer ZIP
Download now
SHA-256: 5c345c32a38f871d46ea0e320d503bb66c5d5a1d114d7431c1b78583dc272e2b
WINDOWS STATUS
The current app does not support Windows.
The downloadable client depends on macOS-only AppKit/SwiftUI, AVFoundation, Accessibility APIs, Apple Translation, and MLX. It cannot be copied or repackaged for Windows.
What works on Windows today?
The source repository includes a Windows 10/11 x64 bootstrap for the official Qwen3-ASR PyTorch backend and a browser microphone demo. It is a local recognition test, not yet a system-wide tray input app.
What is required for a complete Windows release?
A separate .NET 8 desktop client, WASAPI recording, Windows global hotkeys and SendInput, a PyTorch/ONNX model runtime, a local translation replacement, a signed installer, and testing on real Windows 10/11 systems. Practical target: x64, 8 GB RAM minimum, 16 GB recommended, 8–10 GB free disk; an NVIDIA GPU is optional but strongly improves latency.
DETAILS
Before you install
Does CoveType send my voice or transcript to the cloud?
Normal speech recognition and optional AI polishing run locally on your Mac. The first installation downloads models, Apple may download translation language packs, and update checks contact the CoveType release endpoint.
Can CoveType voice typing work offline?
Yes. After the runtime, Qwen models, and any required Apple translation packs have been downloaded, normal transcription and local AI polishing work without an internet connection.
Is CoveType free and open source?
Yes. The CoveType application source is available under GPLv3. Qwen models, Python packages, and Apple frameworks keep their own licenses and terms.
Does it type into any Mac app?
CoveType inserts the result at the current cursor using macOS Accessibility APIs, so it works across editors, browsers, email, chat, documents, and many developer tools.
Is CoveType a full macOS input method?
No. It is a menu-bar voice input utility with a system-wide hold-to-talk shortcut, so you do not need to switch keyboard input sources.
Why does it need Accessibility permission?
macOS requires this permission for global shortcut handling and for inserting text into another active application. You can revoke it at any time in System Settings.
Does local AI polishing support all 30 recognition languages equally?
No. Qwen3.5 is multilingual, but polishing quality is not identical across all languages. Recognition support and rewriting quality are separate; test important language workflows first.