Navo, voice typing for Mac in Arabic and English

Navo is an open source voice keyboard for macOS in the spirit of Wispr Flow. Hold a key, speak naturally, let go, and Navo pastes clean text into whatever app you are typing in. Speech recognition runs on your Mac with your choice of four open models: two for Arabic, Cohere Transcribe Arabic and Audar ASR V1 Turbo, and two multilingual ones that are strongest in English, Whisper Large v3 Turbo and Qwen3-ASR 1.7B. A small local model tidies each transcript, and summaries and rewrites are written on your Mac too, by Gemma 4 (Google) or Llama 3.1 (Meta). After a one-time model download, nothing leaves your computer.

Pastes for you	The text goes straight into the focused app, and your clipboard is put back the way it was
Arabic first	Dialects, Modern Standard, and English terms inside Arabic sentences
Meetings	Live transcript while the call goes on, then a summary with decisions and action items
Transcribe	Voice Memos, WhatsApp voice notes, MP3, M4A, WAV, FLAC, with search and a date filter
Clipboard history	Control + Option + V opens your last copies in any app
Light on memory	Models load when you talk and leave after the idle time you choose
Private	No cloud, no account, no tracking. The engine listens on 127.0.0.1 only

What's new in 0.21

• Live text you can trust. Notetaker and Transcribe write as people speak, then listen to every 20 seconds again as a whole, so the live text is as accurate as a transcription afterwards. Your other meetings stay open while one records.
• Select several, redo them with another engine. Command-click recordings or files and transcribe them again with any engine and language, one after the other, with Stop all.
• One Word file with all of them. Copy or export the texts you select, oldest first, each with its title and date. Arabic stays right to left.
• AI that reads several meetings. Summaries and clean rewrites of several at once, sized to the model's context window so it never stalls, also with the models you already have in LM Studio.
• Measured and quiet. A Performance page (speed, first word, tokens in and out), a problems log, notifications when transcriptions are done, and an emoji of your choice instead of the recording dot in the menu bar.

Every change is in CHANGELOG.md.

Why Navo

The dictation built into your Mac and your iPhone 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.
• Clean text, not a raw transcript: a small model on your Mac adds the punctuation and drops the filler words, without translating or changing 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.

	Built-in dictation (Mac, iPhone)	Paid dictation apps	Navo
Speech model	One general recognizer for everyone	Modern models, often in the cloud	Recent models for your language: Audar or Cohere for Arabic, Whisper or Qwen3 for English
Arabic dialects and mixed sentences	Hit and miss	Depends on the app	Built for it
Clean text	Types what it hears, as it hears it	Yes	Yes, on your Mac
Meetings, with live text and a summary	No	Some, often another app or plan	Included
Hours-long recordings and voice notes	No	Some	Included, as many hours as you like
Where your voice goes	Depends on the language and settings	Often to their servers	Nowhere. It stays on your Mac
Price	Free with your device	About $8 to $15 a month	Free and open source

For creators and long recordings. 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; the text appears piece by piece with times next to every part, to find a moment, write show notes or check a quote; and one click turns it into a summary or a clean rewrite. Notetaker does the same live, for a call or a session while it happens.

Features

• Push to talk anywhere: hold Right ⌥ (configurable: Right ⌘, Right ⌃, fn), release to insert. Double-tap for hands-free, tap again to finish, Esc to cancel.
• See the words while you talk: the text appears in a small card above the Flow Bar as you speak, the finished phrases bright and the one you are saying dimmer. Each phrase is transcribed when you pause, so a long dictation is in almost as soon as you let go. Short ones are still transcribed again in one go, so nothing is lost in accuracy. It can be turned off in Settings.
• Arabic first: MSA, dialects and Arabic-English code-switching. Cleanup keeps your dialect and never translates.
• Switch the language in a second: press Control + Option + L in any app (or record your own shortcut in Settings), use the switch on Home or the AR / EN button on the Flow Bar, even while you talk. Cohere takes Arabic or English; Audar, Whisper and Qwen3 also have Auto, where the model detects the language itself (Audar always picks up the Arabic dialect from the audio).
• Four speech engines, two on at a time: Cohere Transcribe Arabic and Audar ASR V1 Turbo for Arabic dialects, Whisper Large v3 Turbo (100 languages, fast, 1.6 GB) and Qwen3-ASR 1.7B (30 languages) for English and other languages. Pick one for dictation and keep one more on for meetings, files or other apps through the API. Each model takes 1.6 to 4.7 GB while loaded, so a third can be turned on once one of the two is off.
• Compare engines: run any recording or audio file through the two engines that are on and see both transcripts side by side, in the app or with one API call.
• Notetaker, for meetings, with live text: record a Google Meet, Zoom or any call with what your Mac plays, your microphone, or both. The text appears while the meeting goes on: every minute or so (30 seconds to 5 minutes, your choice) the newest piece is cut at a pause and transcribed by the engine you choose. Pause and resume; the audio is saved to disk in parts every 3 or 5 minutes, so a crash loses at most the last part, and becomes one recording when you stop.
• Cuts that never split a word: long audio is only ever cut inside a real pause. Navo looks up to 10 seconds before and after each target point for a stretch where nobody speaks (at least 300 ms, measured against the background noise around it), prefers the longest and cleanest one, and cuts in its middle. Silent pieces are not sent at all, so the models cannot invent text for them. The engine uses the same method to fit pieces into each model.
• Find any recording: Notetaker and Transcribe have a search field that looks in titles and in the text itself (Arabic without caring about diacritics or hamza forms), shows the line where the words were found, and a date filter: today, yesterday, the last 7 or 30 days, or one day picked from a calendar.

How it works

Hold key ─► AVAudioEngine (16 kHz mono) ─► WAV
                                            │
                     Navo engine on 127.0.0.1:7861 (Python, started by the app)
                                            │
          Cohere, Audar, Whisper or Qwen3-ASR (MLX) ─► raw transcript
                                            │
          Local LLM cleanup (Qwen3 4B via MLX) ─► clean text
                                            │
          Clipboard + ⌘V into the focused app ─► SQLite history

AI button ─► the speech models leave memory ─► Gemma 4 or Llama 3.1 (MLX) ─► summary or rewrite

The engine is two processes: a small gateway that always listens on 7861, and a models process that the gateway starts when a model is needed and stops after the idle time. Navo wakes it the moment you start talking, so the model loads while you speak.

The macOS app is native Swift (SwiftUI + AppKit). The engine is a small FastAPI server with OpenAI-compatible endpoints (/cohere/v1/audio/transcriptions, /audar/v1/audio/transcriptions, /whisper/v1/audio/transcriptions, /qwen3/v1/audio/transcriptions, /v1/audio/compare, /gemma/v1/chat/completions, /llama/v1/chat/completions, /v1/chat/completions, /health), so you can also point cleanup at Ollama or LM Studio.

Install

The quick way, one line in Terminal (press Command + Space, type Terminal, press Return, then paste):

curl -fsSL https://devehab.github.io/STT-NAVO/install.sh | bash

It downloads Navo, puts it in Applications and opens it. macOS shows no warning, because a file that comes through curl is not put in quarantine. The script is short, read it first if you like: docs/install.sh. The same line updates Navo later.

Or with the disk image:

1. Download Navo.dmg, open it and drag Navo into Applications.
2. Open Navo from Applications. Navo is not signed with a paid Apple developer account, so macOS says it could not verify it: click Done, open System Settings > Privacy & Security, scroll down and click Open Anyway next to Navo, then confirm with your password or Touch ID. You do this once. (A build signed with an Apple Developer ID and notarized opens with no warning at all, see Make the download.)

Either way, Navo opens on a welcome screen with three steps:

1. Download and set up: Navo brings its own Python, MLX, the Audar speech model and the small cleanup model, about 7 GB, one time, then works offline. No account, no token, and no developer tools (Xcode) are needed.
2. Allow the microphone.
3. Allow Accessibility, so Navo can see the dictation key and paste the text for you.

Then click into any text field, hold Right ⌥, talk, and let go.

More models are one click away in Settings: Whisper and Qwen3 for speech, Gemma or Llama for summaries (5.2 and 4.5 GB). Only Cohere needs an account: accept its terms at huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026, create a Read token at huggingface.co/settings/tokens, paste it in Settings and click Download next to Cohere. Models on this Mac in Settings lists all seven models with their size and whether each is downloaded, and deletes the ones you no longer use (for good, with a confirmation) to free the space.

To update, run the line again, or drag the new version into Applications and replace the old one. Your history, recordings and models stay. If the new version is signed differently, macOS asks for Accessibility and Microphone once more.

Requirements

• Apple Silicon Mac (M1 or newer), macOS 14 or later. 8 GB RAM at least (one speech model, Whisper is the lightest), 16 GB recommended, 24 GB or more to keep two speech engines loaded at once
• About 8 GB free disk space, plus the size of each other model you download. Speech: Whisper about 1.6 GB, Qwen3 about 4.1 GB, Cohere about 4.1 GB, Audar about 4.7 GB. AI writing: Gemma about 5.2 GB, Llama about 4.5 GB
• An internet connection for the one-time engine install
• To build from source: Xcode or the Command Line Tools (xcode-select --install)

Build from source

./scripts/run.sh

run.sh builds Navo.app, installs it to /Applications (replacing a copy an earlier run left in ~/Applications), installs the local engine on the first run (uv, Python 3.12, MLX, PyTorch, the speech model and the cleanup model), and launches Navo. On the first run it asks for your Hugging Face token once (or put HF_TOKEN=... in a .env file, see .env.example). Then:

1. Allow Microphone and Accessibility when Navo asks (Settings shows both).
2. Wait for the sidebar to say Local engine: Ready.
3. Click into any text field, hold Right ⌥, talk, release.

You can also install or reinstall the engine from Navo > Settings > Speech engines. To add Audar, Whisper or Qwen3, click Download next to it there, then Use for dictation if you want it to transcribe what you dictate. Two engines can be on at once: turn one off to turn another on.

Make the download

bash scripts/package.sh

builds Navo and makes dist/Navo-<version>.dmg (with its SHA-256 next to it): a disk image with Navo, a link to Applications and a window that shows people to drag Navo in. Share that file; it holds everything, including the engine installer. macOS may ask once to let Terminal control Finder, which lays out the window.

Without an Apple Developer ID, people allow Navo once on the first open (step 2 of Install); the disk image window says how. With one (Apple Developer Program), the download opens on any Mac with no warning:

# once: store your notary login in the keychain
xcrun notarytool store-credentials navo --apple-id [email protected] --team-id TEAMID
# every release
NAVO_SIGN_IDENTITY="Developer ID Application: Your Name (TEAMID)" NAVO_NOTARY_PROFILE=navo \
  bash scripts/package.sh

This signs Navo with the hardened runtime, sends the disk image to Apple for notarization, and attaches the result so it opens even offline.

Use the engine from other tools

The local engine is an OpenAI-compatible HTTP API on http://127.0.0.1:7861: send audio, get text back. Scripts, plugins and agents can use it while Navo is running.

Engine	Base URL	Documentation
Cohere Transcribe Arabic	http://127.0.0.1:7861/cohere/v1	docs/API-Cohere.md
Audar ASR V1 Turbo	http://127.0.0.1:7861/audar/v1	docs/API-Audar.md
Whisper Large v3 Turbo	http://127.0.0.1:7861/whisper/v1	docs/API-Whisper.md
Qwen3-ASR 1.7B	http://127.0.0.1:7861/qwen3/v1	docs/API-Qwen3.md
Gemma 4 E4B (language model)	http://127.0.0.1:7861/gemma/v1	docs/API-LLM.md
Llama 3.1 8B (language model)	http://127.0.0.1:7861/llama/v1	docs/API-LLM.md

docs/API.md covers what they share: turning engines on and off (two at most at once, a third answers 409), the languages each takes, POST /v1/audio/compare for the transcripts of one file side by side, the cleanup LLM and agent tools. docs/API-LLM.md covers the language models: the system prompt, the text and the response, streaming, the memory rules (asleep until a request wakes them, never loaded with a speech model, at most 6 GB) and long texts. examples/api has ready clients in Python (transcribe.py, compare.py, chat.py), Node and the OpenAI SDK.

Development

bash scripts/build-app.sh debug    # build/Navo.app with swiftc (log in build/build.log)
open Package.swift                 # or work in Xcode through the Swift package

cd engine && python -m pytest -q   # engine tests, no model needed

bash scripts/demo.sh               # open Navo on made-up sample data, for screenshots and demos
bash scripts/publish.sh            # commit and push to GitHub, and turn on the website
bash scripts/release.sh            # build the DMG and upload it as a GitHub release
bash scripts/ship.sh               # all of it for a new version: build, push, DMG and release

demo.sh starts Navo with --demo: the history, the recordings and the clipboard come from a folder of sample data (~/Library/Application Support/Navo/Demo, filled again on every start), your own data is not shown or touched, and the clipboard is not read. Quit Navo and open it normally to go back.

The website is one file, docs/index.html, in English and Arabic (it follows the visitor's browser language), with its fonts in docs/fonts and its pictures in docs/screenshots. It is published with GitHub Pages from the docs folder at devehab.github.io/STT-NAVO. The download link and the repository address are in the NAVO block at the top of its script.

Run the engine by hand (for example to try another model):

PYTHONPATH=engine ~/Library/Application\ Support/Navo/engine/venv/bin/python -m navo_engine --port 7861 --engines cohere,audar
curl -F [email protected] -F language=ar http://127.0.0.1:7861/audar/v1/audio/transcriptions

If you have an Apple Development certificate, build-app.sh signs with it so macOS keeps the Accessibility permission across rebuilds. With ad-hoc signing, remove Navo from System Settings > Privacy & Security > Accessibility and add it again after a rebuild.

Project layout

Sources/Navo/
  App/        AppDelegate (menu bar, lifecycle), MenuBarActivity (transcription progress in the menu bar), AppSettings, DictationController (the pipeline)
  Audio/      AudioRecorder (dictation), microphone and Mac audio capture (Core Audio tap), WAV files
  Input/      HotkeyManager (push-to-talk, Esc), GlobalShortcut (language switch), TextInjector (paste, clipboard restore)
  Engine/     SpeechEngine (Cohere, Audar, Whisper, Qwen3), WritingModel (Gemma, Llama), LocalEngineManager
              (install, download, launch, engines on and off), HTTP clients, CleanupService
  AI/         AIWriter (summaries and rewrites with the local language models), the prompts, long texts in parts
  FlowBar/    Floating edge bubble: panel, layout, SwiftUI views
  Hub/        Main window: Home (stats, history), Notetaker, Transcribe, Clipboard, Dictionary, Settings (with the
              Setup list and the models table), Compare engines, the AI panel
  Sessions/   Meeting recorder (parts on disk, mixing, live pieces), SmartCut and PieceCutter (cuts at pauses),
              audio decoding, pasted and dropped audio, SessionStore (transcription jobs), SessionFilter
              (search and dates)
  Clipboard/  ClipboardStore (clipboard history), QuickPaste (the Spotlight style panel of recent copies)
  Storage/    SQLite database, HistoryStore
  Support/    Paths, Arabic-aware text tools
Extensions/NavoShare/  "Navo" in the Share menu (a sandboxed share extension that hands audio to the app)
engine/
  navo_engine/  FastAPI server, engines (Cohere, Audar, Whisper, Qwen3), ASR backends (MLX, transformers), splitting at
                pauses, language models (Gemma, Llama) and their memory rules, cleanup LLM, model download
  setup-engine.sh, requirements.txt, tests/
scripts/      build-app.sh, run.sh, package.sh (the DMG download), demo.sh (sample data)

Data and privacy

Everything lives in ~/Library/Application Support/Navo: navo.sqlite (history, dictionary, meetings and files, clipboard history), Audio/ (dictation recordings), Sessions/ (meetings and imported files), Inbox/ (audio from the Share menu, moved into Transcribe right away), engine/ (Python environment), Logs/engine.log, Logs/problems.log (what failed without stopping Navo, also in Settings > Performance). Models are in the Hugging Face cache (~/.cache/huggingface). The engine listens on 127.0.0.1 only and loads models with HF_HUB_OFFLINE=1.

• Recordings are kept as WAV files by default so you can replay or re-transcribe them. Turn off Keep recordings and each one is erased right after transcription, or set Keep recordings for to a number of days.
• Deleting is permanent. Recordings are removed from disk (not moved to the Trash). The database runs with secure_delete, so deleted text is overwritten, and it is compacted after every deletion.
• Memory: Settings > Speech engines > Free memory when idle (2 minutes to 1 hour, or never) and Free memory now. Asleep, the engine uses a few dozen MB; the first dictation after a sleep takes a few seconds longer.
• Delete history (Settings) removes everything except the last N days, only the last N days, or everything, for text and recordings or recordings only. Keep text for does the same automatically every hour.
• Clipboard history keeps plain text only, for the time you choose (starred items until you delete them), at most the newest 5,000 items and 20 million characters. Anything marked concealed or transient (password managers do this) and copies from password manager apps are skipped. Deleting is permanent, like everything else.
• AI writing stays on this Mac too. Summaries and rewrites are written by Gemma or Llama inside the local engine. No text, no audio and no key goes anywhere, and it works with Wi-Fi off. Answers are kept in memory while Navo runs and are not saved. (Earlier versions used Google Gemini for this; a key saved back then is erased from your keychain the first time this version runs.)
• The only time Navo uses the internet is to download the engine and the models you choose, once, from PyPI and Hugging Face.

Verify it runs locally

• Turn Wi-Fi off and dictate. It still works.
• Settings > Proof it runs on this Mac shows the device (Apple GPU via MLX), the address (127.0.0.1 only), offline mode, engine memory, model sizes on disk and how long the last transcription took. Every history entry shows On this Mac, MLX and its processing time.
• From the Terminal:

curl -s http://127.0.0.1:7861/health | python3 -m json.tool
ps -o pid,rss,%cpu,command -p $(pgrep -d, -f navo_engine)   # the gateway, plus the models process while awake
lsof -nP -iTCP -sTCP:LISTEN | grep 7861

Troubleshooting

Problem	Fix
The key does nothing	Allow Accessibility for Navo, then wait two seconds (Navo re-arms the key automatically).
Text is copied but not pasted	Same: Accessibility is required to send ⌘V. Password fields are never auto-pasted.
fn opens the emoji picker	System Settings > Keyboard > "Press 🌐 key to" > Do Nothing, or use Right ⌥.
Install says access refused	Accept the model terms on Hugging Face with the account that owns the token.
Engine shows Error	Settings > Engine log. For Cohere, MLX failures fall back to PyTorch automatically.
The AI panel says a model is not downloaded	Click Download there, or in Settings > AI writing. Gemma is 5.2 GB, Llama 4.5 GB, no token needed.
Gemma says the model type is not supported	The engine was installed before Gemma 4 existed. Settings > Speech engines > Reinstall / update; the models you have are kept.
Dictation is slower right after a summary	The speech model had stepped out of memory for the language model and loads again, a few seconds. Closing the AI panel frees the language model at once.
Audar says Not downloaded	Settings > Speech engines > Download next to Audar. No token is needed.
The Mac feels slow with two engines on	Lower Free memory when idle, or turn off the engine you do not use for dictation, in Settings > Speech engines.
An engine's switch is greyed out	Two engines are on already, and that is the most at once. Turn one of them off, then turn this one on.
Notetaker shows no sound from the Mac	System Settings > Privacy & Security > Screen & System Audio Recording > System Audio Recording Only: allow Navo, then start the recording again. Needs macOS 14.4 or later.
The first dictation after a pause is slower	The models were asleep and loaded again. Raise Free memory when idle, or set it to Never.
Navo is not in the Share menu	Run ./scripts/run.sh once (it registers the extension), then System Settings > General > Login Items & Extensions > Sharing (on macOS 14: Privacy & Security > Extensions > Sharing): turn Navo on. In the Share menu, Edit Extensions shows it too.
Paste says there is no audio	The message lists what the clipboard holds. Use Share > Navo in Voice Memos instead, or drag the recording to Finder and then onto Transcribe.

Credits

• Cohere Transcribe Arabic by Cohere Labs, Apache 2.0
• Audar ASR V1 Turbo by AudarAI, AudarAI Community License v1.0
• Whisper Large v3 Turbo by OpenAI, MIT, in mlx-community's MLX build
• Qwen3-ASR 1.7B by the Qwen team at Alibaba, Apache 2.0, in mlx-community's MLX build
• Gemma 4 E4B by Google, Apache 2.0, in mlx-community's 4 bit MLX build
• Llama 3.1 8B Instruct by Meta, Llama 3.1 Community License, in mlx-community's 4 bit MLX build
• mlx-audio and mlx-lm for Apple Silicon inference
• Inspired by Wispr Flow

MIT License.