Tap. Speak. It's typed.
The whole loop takes a few seconds, in any app you're already in.
Listening
Tap right ⌘ anywhere. The HUD floats above whatever you're doing — hold-to-talk and Esc-to-cancel work too.
Transcribing
Tap again to finish. The fast engine runs at ~12× realtime, so a spoken sentence comes back in about a second.
Inserted
The text lands at your cursor — native apps, browsers, Electron, terminals. No window switching, no copy-paste.
Small app. Serious engine room.
A ~2 MB native menu-bar app in front of two always-warm speech engines — all of it running on your Mac.
One key is the whole interface.
Tap to talk, tap to finish. Hold to talk instead, Esc to cancel — or remap it to any key you like. It works wherever your cursor is.
Two engines, both always warm.
The fast model answers dictation instantly; the accurate one handles the careful second passes. Switching is instant routing — never a model reload, never a spinner.
Sharpen: a second opinion, one tap.
Your audio stays on your Mac during the session, so one tap re-transcribes it on the accurate engine and swaps in the better text. Cloud tools that discard your audio can't offer this.
中英混说, without the mess.
Built on Qwen3-ASR, which reports ~1.48% character error rate on Chinese benchmarks where Whisper-family models sit near 5%. Your names and jargon go in a Vocabulary that biases recognition everywhere.
Enhance: cleanup that never blocks.
An optional local 4B model strips fillers and fixes punctuation in ~0.7 s, under a hard 8 s budget — if cleanup is ever slow, Mynah types the raw transcript instead of making you wait.
A library that's actually yours.
Every dictation is stored on your disk — audio plus raw, enhanced, and sharpened text — with CJK-aware full-text search. Delete anything, anytime.
Meetings, transcribed where they happened: on your Mac.Coming in an update
Record long-form from the mic. When you stop, the accurate engine transcribes the whole recording, speaker labels follow as a post-pass, and the local LLM can clean it up or summarize it. No bots joining your calls — and no audio uploaded, ever.
Local by architecture, not by promise.
Mynah's speech engines listen only on your Mac's loopback address. There is no server that could receive your audio — nothing to trust, nothing to breach.
No account. No telemetry. No device IDs.
Nothing to sign up for, nothing phoning home. The only usage signal that exists is this site's anonymous update log — date, app version, macOS version, chip architecture. IPs are dropped, and you can switch update checks off.
A badge that can't lie.
Mynah shows a live privacy badge. If you ever point the optional Enhance step at a remote model, it flips to an amber open lock that says so — speech recognition stays local regardless. You always know where bytes go.
Your data has an exit.
Transcripts and audio live in your local library; delete them anytime. The built-in uninstaller removes the background service and the models — and, if you choose, every last transcript.
Get Mynah for your Mac.
Mynah for macOS
Download MynahRequirements
- Apple Silicon Mac (M1+)
- macOS 14 or newer
- 25 GB free disk space during setup
- 16 GB+ memory (24 GB+ recommended)
What happens next
-
Unzip, drag into Applications, open. Notarized by Apple — no Terminal, no Homebrew, no Xcode.
-
The setup wizard does the heavy lifting. One click installs the local engine (~400 MB) as a background service — that's what the one password prompt is for — then downloads the speech models (up to ~11 GB on disk), checksum-verified and resumable.
-
Grant the mic and Accessibility, tap right ⌘, speak. Accessibility is what lets text land at your cursor — without it, results go to the clipboard for ⌘V.
Served over HTTPS — the engine and models are SHA-256-verified, the app is signed and notarized by Apple · Uninstalls cleanly, models included
Questions & answers.
Is Mynah free?
Free while it's in beta — no account, nothing locked. Mynah will be a paid app at 1.0; right now, your feedback is the price.
Why is setup such a big download?
That's the speech models: the fast and accurate engines plus the local cleanup LLM. They download once — verified and resumable — live in your home folder, and are the reason nothing needs the cloud afterwards. You can re-download or unload them anytime in Settings.
Does it work offline?
Yes. After setup, dictation, cleanup, and search are fully offline. The network is used only for the initial download and the optional update check.
Will it run on an Intel Mac?
No — Mynah is built on MLX, Apple's machine-learning framework for Apple Silicon, so it needs an M1 or later. That constraint is also what makes two always-warm engines practical on a laptop.
What does “always warm” cost?
Memory, not battery: the resident speech engines take about 2.1 GB of RAM and idle at ~0% CPU, so they don't drain anything while waiting. 16 GB Macs run everything; if memory is tight, Settings can unload the bigger models and keep just dictation warm.
What exactly leaves my Mac?
Audio and transcripts: nothing, ever. If update checks are on, the request carries only the app version, macOS version, and chip architecture — server logs keep those plus the date, and IPs are dropped. If you deliberately point Enhance at a remote API, only the text of that one request goes out, with the badge showing an amber open lock the whole time.
How do updates work?
Mynah checks this site and updates in place — the standard Sparkle mechanism used by most independent Mac apps. Engine and model updates flow through the same wizard, verified like the first install. You can turn checks off in Settings.
Which models power it, and under what licenses?
Qwen3-ASR 0.6B and 1.7B, a Qwen3 forced aligner, and a Qwen3.5-4B cleanup model — all Apache-2.0, all running only on your Mac. (The meetings update will add pyannote speaker diarization, CC-BY-4.0.)
How do I uninstall?
With the built-in uninstaller: it stops and removes the background service and the downloaded models, and asks whether to keep or delete your library. The app itself you drag to the Trash, like any Mac app.
Found a bug? Want something?
Mynah is built by one person, and this form is read by that person. Every report matters at this stage — thank you.
Got it — thank you.
Every report gets read. If you left an email, you'll hear back.