Haki (حكي): push-to-talk dictation for the Mac menu bar, tuned for Arabic

What is Haki?

Haki (حكي, a Levantine Arabic word for talking) is a lightweight native macOS menu-bar app for speech-to-text dictation. Click into any text field in any app, hold the push-to-talk hotkey, speak, and let go. Haki transcribes what you said and pastes it at the cursor. Its tagline in Arabic is "احكي، وحكي بيكتبها", roughly "you talk, Haki writes it".

It is designed especially for natural Arabic and Levantine speech, with support for English, German, French and automatic language detection. It has no local speech model and no Python or Node.js runtime: the app is native Swift, and the transcription is done by OpenAI.

How dictation works

1. Focus a text field in any application, such as Notes, TextEdit, a browser, a chat app or an editor. 2. Press and hold ⌥ Option + Space and speak. 3. Release the keys. Haki sends the recording to OpenAI, receives the text and pastes it at your cursor.

Haki saves your clipboard before pasting and restores it afterwards, and it does this without stealing focus from the app you are typing in. A small floating indicator shows whether it is recording, transcribing or done. Accidental taps shorter than about a third of a second are discarded locally.

Languages and models

• Languages: Arabic (with optimization for Levantine and Syrian speech), English, German, French, or automatic detection.
• Models: gpt-transcribe is the default and is aimed at fast, low-latency daily voice typing. gpt-4o-transcribe is the higher-accuracy option for complex phrases, dialects and technical terms. You can switch from Settings or from the menu bar.
• Microphones: pick the built-in microphone, a USB interface or AirPods. Haki handles reconnection through CoreAudio and can prefer AirPods when they connect. A live level meter lets you test the input.

Privacy: what stays on your Mac and what goes to OpenAI

Be clear about the trade-off: your voice recordings are sent to OpenAI. According to the README they go directly from your Mac to OpenAI's transcription endpoint over TLS, with no intermediary servers, so dictation needs an internet connection and your own OpenAI API key, used under OpenAI's terms and pricing.

On your Mac, the README says Haki embeds no analytics, diagnostics, crash reporters or third-party SDKs. Audio is recorded to temporary files and deleted as soon as transcription finishes or fails, so no audio history is kept. Your API key is stored in the macOS Keychain and not in preferences, configuration files or logs.

Permissions

• Microphone: to capture your voice, and only while you hold the hotkey or start dictation from the menu.
• Accessibility: for the global push-to-talk hotkey and to insert text by simulating ⌘V. According to the README, Input Monitoring is not required separately.

Set it up

1. Open Haki. Allow Microphone access when macOS asks, and enable Haki under System Settings, Privacy & Security, Accessibility. 2. Open Settings from the menu-bar icon or with ⌘,. In the Transcription tab paste your OpenAI API key, choose Save, then Test API Key. 3. Choose a model, and in the Audio tab choose your microphone.

Dictate from your iPhone with AirMouse

AirMouse is a separate open-source app that turns an iPhone into a trackpad for your Mac over your local Wi-Fi. Its voice tab records on the phone and hands the audio to Haki, which types the text on your Mac. Haki has to be installed and running, and it uses your normal API key, model and language settings.

The handoff is simple and open to other senders too. A recording (m4a, mp4, webm, wav, mp3 or ogg, up to 25 MB) is written to ~/Library/Caches/Haki/Inbox/, then the URL haki://transcribe?file=<file name> is opened. Haki transcribes the file, pastes the text and deletes the file. Only plain file names inside the inbox are accepted, so paths, hidden files and symlinks are rejected.

Build it

Haki needs macOS 13 Ventura or later, on Apple silicon or Intel, and the Xcode Command Line Tools with Swift 6.0 or newer. Clone the repository from GitHub, run swift test, then ./scripts/build_app.sh. The script builds the release binary, assembles Haki.app, signs it with your local Apple Development certificate or an ad-hoc signature, and installs it in /Applications.

Open source

Haki is released under the MIT license. The contribution guide asks for no local models, no Python, no Node.js and no Electron, and for API keys never to be committed. Emad Alhamou, the author of this site, publishes Haki on GitHub. Bug reports and pull requests are welcome, and security concerns go through the repository's security policy.