See the words while you talk
A small card above the Flow Bar shows what you say as you say it. Long dictations are written as you go, so they are in almost as soon as you let go.
Talk instead of typing. In Arabic, in English, or in both within one sentence.
Navo is an open source voice keyboard for macOS. Hold one key, speak naturally, release, and Navo pastes clean text into the app you are working in. Speech recognition, summaries and your history all stay on your Mac.
Try it: hold this key, or Option on your keyboard, to hear a real dictation and watch the words appear.
v0.13.0 October 2026
What is new in this version
Set up in three clicks
A welcome screen walks you through it: download the engine, allow the microphone, allow Accessibility.
No account, no developer tools
The first setup needs no Hugging Face token, and macOS no longer asks to install Xcode tools.
Install with one line
Paste one line in Terminal and Navo downloads, installs and opens, with no warning from macOS.
Hide the Flow Bar
For 10 minutes up to 2 hours, or keep it out of sight and see it only while you dictate.
Free space in one click
Delete any model you no longer use, with a confirmation that says how much space comes back.
Record is now called Notetaker, and Files is now Transcribe. All changes
Why Navo
Your Mac already types what you say. Navo understands it.
The dictation built into your Mac and your phone is free and handy for a short sentence in a quiet room. Real speech is messier: a café, a car, a dialect, an accent, an English word in the middle of an Arabic sentence. That is where one general recognizer, made for everyone, slips. Navo runs recent speech models built for exactly this, and you pick the one that fits the way you speak.
Noise around you
Qwen3-ASR is built and tested for loud, messy audio and for English spoken with many accents. Navo also leaves the silent stretches out, so background hum never turns into words.
Your dialect, not only the textbook
Audar learned from more than 300,000 hours of speech: Gulf, Egyptian, Levantine and Maghrebi Arabic, Modern Standard, and English.
Two languages in one sentence
Say staging or release inside an Arabic sentence and it comes out in English letters, not as a guess in Arabic script.
Clean text, not a raw transcript
A small model on your Mac adds the punctuation and drops the filler words. It never translates and never changes your dialect.
Audar ASR V1 Turbo and Cohere Transcribe Arabic, both in Navo, rank first and second on the Open Universal Arabic ASR Leaderboard, as their makers report, at about 23% and 26% average word error rate. Whisper Large v3, the model many dictation apps are built on, is at about 37% on the same tests.
Navo next to the alternatives
| What you get | Built-in dictationMac and iPhone | Paid dictation appsBy subscription | NavoFree and open source |
|---|---|---|---|
| Speech model | Built-in dictationOne general recognizer for everyone | Paid dictation appsModern models, often in the cloud | NavoRecent models for your language: Audar or Cohere for Arabic, Whisper or Qwen3 for English |
| Arabic dialects and mixed sentences | Built-in dictationHit and miss | Paid dictation appsDepends on the app | NavoBuilt for it |
| Clean text | Built-in dictationTypes what it hears, as it hears it | Paid dictation appsYes | NavoYes, on your Mac |
| Meetings, with live text and a summary | Built-in dictationNo | Paid dictation appsSome, often another app or plan | NavoIncluded |
| Hours-long recordings and voice notes | Built-in dictationNo | Paid dictation appsSome | NavoIncluded, as many hours as you like |
| Where your voice goes | Built-in dictationDepends on the language and settings | Paid dictation appsOften to their servers | NavoNowhere. It stays on your Mac |
| Price | Built-in dictationFree with your device | Paid dictation appsAbout $8 to $15 a month | NavoFree |
Hours of audio, no bill by the minute
For podcasters, YouTubers, journalists, teachers and students. Most transcription services charge by the minute or by a monthly plan. With Navo, a two-hour interview, a lecture or a pile of voice notes becomes text on your Mac, as often as you like, for free.
- Long audio is cut only inside a pause, so no word is split, and the text appears piece by piece while you watch.
- Times next to every part, to find a moment, write show notes or check a quote.
- One click turns it into a summary, the main points or a clean rewrite, in English or Arabic.
- Notetaker records a call or a live session while the text appears as it goes.
Three moves, in any app
-
Hold
Hold Right Option wherever you can type. The Flow Bar opens with a live waveform, so you see that Navo hears you.
-
Speak
Talk the way you normally do. Arabic dialects, English, and switching between them in the middle of a sentence all work.
-
Let go
Navo tidies the text and pastes it where your cursor is. You do not copy, switch windows or press anything. Double tap the key to talk hands free, Esc cancels.
What Navo does for you
Dictation is where it starts. Navo also writes down your meetings, turns recordings into text, writes the summary and remembers what you copied. None of it leaves your Mac.
It pastes the text for you
Let go of the key and the sentence is already in your email. Navo puts the text on the clipboard, presses Command V in the app you are in, then puts back whatever you had copied before. You never switch windows and you never paste.
- You see the words in a small card while you talk, and a long dictation is written as you go, so it is in almost as soon as you let go.
- Works wherever a text field has focus: Mail, Slack, the browser, your code editor.
- First a small model on your Mac fixes the punctuation and drops the filler words. It never translates and never changes your dialect.
- No text field in focus? The text waits on the clipboard and Navo tells you. Password fields are never filled.
- Every dictation stays on your Mac with its recording: search it, replay it, retry it.
Arabic and English, one shortcut apart
Press Control Option L in any app, or click the AR / EN chip on the Flow Bar, even in the middle of a sentence.
- Arabic the way people speak it: dialects, not only Modern Standard.
- Say an English term inside an Arabic sentence and it stays in English letters.
- In Auto, the model works out the language by itself.
Your meeting, written down while it happens
Record a Google Meet, a Zoom call or anything your Mac plays, your microphone, or both. Every minute or so the newest part turns into text, so you can read along while people are still talking.
- The audio is only ever cut inside a pause, so no word is split in two.
- Pause and resume. The recording is saved in parts, so a crash loses almost nothing.
- The menu bar shows a red dot and the time while you record.
The summary is ready before you close the laptop
One click on AI turns an hour of talk into the main points, the decisions and the action items, with the names and the numbers. It also rewrites any text as a formal email, a friendly email, a text message or a clean draft, in English or Arabic, whatever language you spoke.
- Gemma 4 or Llama 3.1 writes it on your Mac, and you watch the answer appear.
- No account, no API key. It works with Wi-Fi off.
- Long meetings are read in parts, so an hour of talk fits.
Any recording becomes text you can search
Drop in Voice Memos, WhatsApp voice notes, MP3, M4A, WAV or FLAC. Navo writes them out piece by piece, with times if you want them, ready to copy or export.
- Search looks in the titles and inside the text. Arabic matches without diacritics or hamza forms.
- Filter by date: today, yesterday, the last 7 or 30 days, or one day from the calendar.
- Send audio in with Share in Voice Memos, a paste, a drop, or Open With in Finder.
Everything you copied, one shortcut away
Press Control Option V in any app. A small panel opens over your work, like Spotlight, with your last ten copies. Type to search everything Navo kept, press Return, and it is pasted right where you were.
- Copied something on your iPhone? With Universal Clipboard on, it lands here too.
- Star what you want to keep forever. The rest leaves after the time you choose, from 1 day to 2 months.
- Copies from password managers are never kept.
Fast when you talk, gone when you stop
The models run on your Mac's GPU through MLX, Apple's framework for Apple Silicon. Navo wakes the engine the moment you press the key, so the model loads while you are still speaking. When you are done, it gives the memory back.
- You choose the idle time: 2, 5, 10, 30 or 60 minutes.
- Asleep, the engine takes a few dozen MB instead of several GB.
- A speech model and a language model are never in memory together, and a language model never takes more than 6 GB.
Small things that add up
No dictation is ever lost
If transcription fails, the recording stays in your history with a Retry button.
Your own dictionary
Teach it names and terms, and set exact replacements for what it should write.
Two engines, side by side
Run one recording through two speech models and compare what each one wrote.
Progress in the menu bar
The icon fills like a pie while a long file is transcribed, then turns into a green check.
No invented words
Silent stretches are never sent to the model, so it cannot make up text for them.
Delete means delete
Recordings are erased from disk, not moved to the Trash, and deleted text is overwritten.
Your keys, your way
Dictate with Right Option, Right Command, Right Control or fn, and record your own shortcuts.
A local API
Your scripts and other apps can use the same models through an OpenAI compatible API.
100% private. Everything runs on your Mac.
No cloud, no account, no tracking. The engine listens on 127.0.0.1 only and loads its models in offline mode. The internet is used once, to download the models you choose. Turn Wi-Fi off and dictate: it still works.
- Your voice and your text never leave the device.
- Deleting is permanent: recordings are erased from disk and deleted text is overwritten.
- Settings shows the proof: the device, the address, the memory in use and how long each transcription took.
- The code is open. Read it, build it, change it.
The models
All of them are open models, running through MLX on the Apple Silicon GPU. Download only the ones you need.
Speech to text
| Model | Best for | Download | Hugging Face token |
|---|---|---|---|
| Audar ASR V1 TurboAudarAI | Arabic dialects, and Arabic with English mixed in. Installed by default. | 4.7 GB | Not needed |
| Cohere Transcribe ArabicCohere Labs | Arabic, Modern Standard and dialects, and English. | 4.1 GB | Free token needed |
| Whisper Large v3 TurboOpenAI | English and about 100 languages. The lightest and the fastest to load. | 1.6 GB | Not needed |
| Qwen3-ASR 1.7BQwen, Alibaba | English and 30 languages. | 4.1 GB | Not needed |
Two speech models can be on at the same time: one for dictation, and one more for meetings, files or the API.
Language models
| Model | Best for | Download | Hugging Face token |
|---|---|---|---|
| Gemma 4 E4BGoogle | Summaries and rewrites. The better choice for Arabic. | 5.2 GB | Not needed |
| Llama 3.1 8BMeta | Summaries and rewrites. Strong in English. | 4.5 GB | Not needed |
| Qwen3 4BQwen, Alibaba | Tidies each dictation: punctuation and filler words. Installed by default. | 2.3 GB | Not needed |
Gemma and Llama are optional. They are asleep until you ask for a summary, and they leave memory again when you close the panel.
What you need
- Mac
- Apple Silicon (M1 or newer). Intel Macs are not supported.
- System
- macOS 14 Sonoma or later. Recording what the Mac plays in a meeting needs macOS 14.4 or later.
- Memory
- 8 GB at least, 16 GB recommended. With 8 GB, keep one speech model on (Whisper is the lightest). Summaries are comfortable with 16 GB.
- Disk
- About 8 GB free for the first setup, plus the size of each extra model, from 1.6 to 5.2 GB.
- Internet
- Only for the first download. After that Navo works offline.
- Hugging Face token
- Only for the Cohere model: a free account and a Read token. Every other model downloads without an account.
- Permissions
- Microphone, and Accessibility so Navo can paste the text. System Audio Recording if you record meetings.
Install and first setup
Three clicks, then a one time download that runs by itself.
- Install Navo with the one line in Terminal, or with the DMG file. Navo opens on a welcome screen with three steps.
- Click Download and set up. Navo brings its own Python, MLX, a speech model and the small cleanup model, about 7 GB, one time. No account, no token and no developer tools are needed.
- Click Allow for the Microphone.
- Click Open Settings for Accessibility and turn Navo on in the list. It lets Navo paste the text for you.
- When the download finishes, click into any text field, hold Right Option, talk, and let go.
Later, in Settings, you can add more models: Whisper and Qwen3 with one click, Gemma or Llama for summaries, and Cohere, the only one that needs a free Hugging Face token.
Shortcuts
- Hold Right Option
- Dictate while the key is down
- Double tap Right Option
- Dictate hands free, tap again to finish
- Esc
- Cancel the dictation
- Control Option L
- Switch between Arabic and English
- Control Option V
- Quick paste from the clipboard history
- Command 1 to 0
- Paste item 1 to 10 from the quick paste panel
The dictation key and both shortcuts can be changed in Settings.
For developers: a local API
The engine is an OpenAI compatible API on http://127.0.0.1:7861. Send audio and get text, or send text to a language model, from any script or tool while Navo is running. No key is needed.
To build from source, run ./scripts/run.sh. The app is native Swift (SwiftUI and AppKit) and the engine is Python (FastAPI and MLX).
# Audio in, text out
curl -F file=@meeting.m4a -F language=ar \
http://127.0.0.1:7861/audar/v1/audio/transcriptions
# Text in, summary out
curl http://127.0.0.1:7861/gemma/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"messages": [
{"role": "system", "content": "Summarize in three points."},
{"role": "user", "content": "..."}]}'
Download Navo
One file. After the first setup everything works without the internet.
Download Navo for MacEasiest Install with one line
Open Terminal (press Command Space, type Terminal, press Return), paste this line and press Return:
curl -fsSL https://devehab.github.io/STT-NAVO/install.sh | bash
Navo downloads, goes into Applications and opens by itself. macOS shows no warning and there is nothing to change in System Settings. The same line updates Navo later.
Or with the DMG file
Navo is free and open source, and it is not signed with a paid Apple developer account. So when it comes from a browser download, macOS blocks it the first time and asks you to allow it yourself. You do this once.
- Open the downloaded Navo.dmg and drag Navo onto Applications.
- Open Navo from Applications (or from Launchpad). macOS says it could not verify the app. Click Done, do not move it to the Trash.
- Open System Settings > Privacy & Security and scroll down to Security. Next to "Navo was blocked", click Open Anyway.
- Click Open Anyway again and confirm with your password or Touch ID. Navo opens, and from now on it opens like any other app.
Do not want to trust a download? The code is open: read it and build Navo yourself with ./scripts/run.sh.