Free dictation for Mac that runs entirely on your machine. Hold one shortcut, say the sentence, let go, and the words land in the field you were already typing in.
Apple SiliconmacOS 14.2+Free, no account
Launch checklist
Landing copy is done. The icons need one more pass.
Calls and transcripts can now be viewed by week and by month with a total for each day, searched by text, and labelled with the names of the people in them. Once you name someone, the app remembers their voice and labels them in later recordings by itself. All of this is computed and stored on your Mac.
New in 0.16.0
The week view places recordings at their time of day, the month view shows how much was recorded each day. Search covers text, notes and names, and the match is highlighted. Click the “Other side” label and give the person a name: it is written into the transcript file, and the voiceprint stays on your Mac. The player shows who spoke when and lets you pick the speed.
Telegram, WhatsApp, Meet, Discord
Press once at the start of a call and once at the end: in the menu bar or on the prompt that appears when a call starts. Your microphone and the sound your Mac plays are recorded as two tracks, so both sides are captured whole. They are cut into replies, recognised on your Mac and merged into one conversation with timecodes. When the sound was kept, the call plays inside the app: press the time of a reply and the recording starts there. You can add a note in your own words to a recording, and it shows in the list instead of a bare date.
From a recorder, a phone, a messenger
A handheld recorder, a phone, a voice message, the sound of a meeting video. Drop the file on the app or leave it in the Inbox folder and it is picked up. What comes out is plain text in paragraphs with no timecodes, and when several people speak the app tries to tell them apart. Imported files are never deleted.
This is a mock-up with made-up recordings. The app works the same way, only with your own conversations and on your Mac.
Switch between List, Week and Month. Under each date is the total recorded that day. Click a recording to open it.
No window to open, no button to find. The app waits in the menu bar and answers to one shortcut.
Put the cursor where you write. The shortcut is yours to change if this one is already taken, and you can switch to press-once mode instead of holding.
Your Mac listens through the model you downloaded. Nothing is streamed to a server, because there is no server of ours to stream to.
The text is inserted at the cursor and stays on the clipboard as a backup. Enter is never pressed for you, so nothing gets sent by accident.
Notes, mail, chats and prompts. The same shortcut everywhere, because the app types into the field you are in instead of asking you to switch apps.
Prototype first, then ask the team what breaks.
Spoken in 4 secondsThanks for the draft, comments coming before Thursday.
Spoken in 5 secondsRunning ten minutes late, order me a coffee.
Spoken in 3 secondsDraft a launch plan from the notes I dictated this morning.
Spoken in 5 secondsIf a field cannot take the text (a password box, a protected input, an unusual editor), the result is already on your clipboard, waiting for one paste.
Speech is recognized on-device by the model you downloaded. There is no account to create, and no server of ours for audio to travel to.
Free · Works offline
Download one model inside the app. It comes from Hugging Face once; after that you can dictate on a plane.
Our pick for everyday dictation. Wider multilingual coverage, and steadier with names and terms the faster model trips on.
The fast model: a short pause between letting go and seeing the text. 25 languages, English and Russian among them.
Both run locally and can be switched in settings. The heavier models are chosen there too: the full Whisper Large v3 and Canary 1B v2. Every model has a card in the app with its pros and cons. Model weights are not bundled into the app, and their licences are listed inside it.
One download and two permissions stand between the DMG and your first dictated sentence.
Open the DMG and drag Bocha Voice into Applications. Launch it from there.
Pick a model, we suggest Whisper Large v3 Turbo, and let it download. Allow the microphone, and Accessibility if you want the text inserted for you.
Choose a shortcut, put the cursor in any text field, hold it and speak. The app walks you through one practice round.
What people ask before installing an app that listens to them.
To download the app, and to download a model the first time. After that recognition happens on your Mac, and dictation keeps working with Wi-Fi off.
The network is needed in three more cases. Once a day the app checks whether a new version is out. The first time a call or a recording you bring in is transcribed, a model that tells voices apart is downloaded. The languages for dictation with translation are downloaded once by macOS itself.
Apple Silicon (M1 and newer) on macOS 14.2 or later. Dictation with translation requires macOS 26 or later. There is no Intel, Windows or phone version. Keep about 630 MB free for the model.
In the text fields of ordinary apps: notes, mail, chats, editors, AI prompt boxes. Automatic insertion needs the Accessibility permission and works while the field you started in is still focused.
If a field cannot take the text, the result waits on the clipboard for you to paste.
This build is signed with the project certificate but has not been notarized by Apple, so Gatekeeper asks before the first launch. The standard way to allow it is System Settings → Privacy & Security, and only if you trust the source.
There are no scripts here that switch off macOS protections.
Yes. No account, no subscription, no payment and no dictation limit. There is no cloud processing to bill for.
Dictation handles English, Russian and many other languages: 25 with Parakeet, more with Whisper. The app interface follows your macOS language and is available in English and Russian.
The recording exists only while recognition runs and is deleted afterwards; after a crash a temporary file may remain. Dictation history is off by default. Settings can turn it on so the app holds the last five, ten or twenty phrases on this Mac and you can get an earlier one back; the list clears with one button. Either way nothing is sent to us. Your settings and downloaded models stay on your Mac.
Transcripts of calls and of recordings you bring are Markdown files in the folder you chose. The sound of a call is deleted after transcription unless you decided to keep it.
The names you give to people are written into the transcript file. Their voiceprints are stored on your Mac and are erased with one button in settings.
Yes, that is what dictation with translation is for. Give it a separate keyboard shortcut in settings and download the languages there. macOS itself translates on your Mac, and the text is not sent anywhere. It requires macOS 26 or later.
You name a person in a transcript yourself, and the app keeps a voiceprint: a set of numbers that tells one voice from another. In later recordings a similar voice gets that name. Speech cannot be rebuilt from a voiceprint. Voiceprints are computed and stored only on your Mac, and the Forget voices button in settings erases them all. Recognition is approximate: check the labels when it matters.
Settings have a personal dictionary. Terms such as Cloudflare or GitHub are given to the model as a hint and corrected by sound after transcription. Exact replacements swap one spelling for another. The dictionary is kept on your Mac.
Not yet: the repository is private while the app is in an early build. The licences of every third-party component and speech model it uses are listed inside the app.
Install it, pick a model, and dictate. Nothing to configure in the cloud, because there is no cloud.
Apple SiliconmacOS 14.2+Free, no account
