HeyVedu

Private, push-to-talk dictation for macOS. Hold Control + Option (or a shortcut you choose), speak, and release to insert polished text into the app you are using.

Speech recognition and the default Vedu Scribe text cleanup run on your Mac. Audio stays in memory, transcripts are not logged, and the default dictation flow does not upload your speech or text. The default flow needs an internet connection only for the initial build and model downloads.

What it does

• Dictates into almost any app that accepts pasted text.
• Starts listening while you hold Control + Option and stops when you release it. You can pick another shortcut, such as fn or the right ⌘ key.
• Tap the shortcut once to dictate hands-free, then press it again to finish.
• Transcribes English speech locally with NVIDIA Parakeet TDT 0.6B v3 through FluidAudio and Core ML.
• Removes fillers and false starts, applies self-corrections, and fixes punctuation and capitalization.
• Offers local cleanup with Vedu Scribe, HeyVedu's own model (the default), or Apple Intelligence.
• Writes names and terms from your personal dictionary exactly as you spell them, and expands spoken snippets into saved text.
• Lets you select any connected microphone and follows device changes.
• Keeps no dictation history and never writes recorded audio to disk. Only the last result stays in memory, for five minutes, so you can paste it again.

For example:

“I want to book a meeting for two, actually no, three PM”
becomes “I want to book a meeting for 3 PM.”

Requirements

• An Apple silicon Mac
• macOS 14 (Sonoma) or later
• Xcode with the macOS 26.4 SDK
• An internet connection for the initial build and first-run model downloads
• About 1.2 GB of free space for the speech and Vedu Scribe models, plus build data

Apple Intelligence cleanup additionally requires macOS 26 or later, a supported Mac with Apple Intelligence enabled, and its model downloaded. Vedu Scribe does not require Apple Intelligence.

Download

Download the latest signed and notarized release from app.heyvedu.com/download, open the disk image, and drag HeyVedu to Applications. The app checks for updates and installs them itself.

Or install it with Homebrew:

brew install --cask rajan9519/tap/heyvedu

Install from source

Open HeyVedu.xcodeproj in Xcode, select the HeyVedu scheme and press ⌘R. Xcode may ask you to trust the pinned MLX build-tool plugin; review the prompt and choose Trust & Enable to continue.

Debug builds are named HeyVedu Dev (bundle ID com.heyvedu.app.dev) and run from Xcode's DerivedData, so they never replace or collide with an installed HeyVedu. They ask for Microphone and Accessibility separately from the installed app.

Debug builds are ad-hoc signed unless you configure a certificate, and macOS forgets an ad-hoc app's Accessibility grant after every rebuild. To keep the grants across rebuilds, sign Debug builds with any certificate you own:

cp Config/DebugSigning.local.xcconfig.example Config/DebugSigning.local.xcconfig
# set CODE_SIGN_IDENTITY and DEVELOPMENT_TEAM; the file is gitignored

To build and launch the same Debug app without Xcode's UI:

./scripts/run.sh

If xcode-select points to Command Line Tools instead of Xcode, set DEVELOPER_DIR for that invocation:

DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer ./scripts/run.sh

To run the unit tests (⌘U in Xcode). They load into the Debug app, which skips permissions and model downloads while tests run:

xcodebuild -project HeyVedu.xcodeproj -scheme HeyVedu -skipPackagePluginValidation test

Build a DMG release

The release script builds an Apple silicon (arm64) DMG for macOS 14 or later, signs the app and disk image with Developer ID, and notarizes and staples both. Intel Macs are not supported.

cp Config/release.env.example Config/release.env
# Fill in your Developer ID identity and Team ID, and store notarization
# credentials in Keychain as described in docs/RELEASING.md.
./scripts/release.sh

For a local packaging test without Apple credentials:

./scripts/release.sh --unsigned

Outputs go to dist/. The unsigned test image is clearly labeled and is not ready for public distribution. See Releasing for complete signing, Apple ID setup, notarization, and installation verification steps.

First launch

HeyVedu opens a short welcome guide that walks through the steps below and a practice dictation. Once you finish it, the HeyVedu window opens, and it opens again each time you launch the app or click its Dock icon. Closing the window leaves HeyVedu running in the menu bar. Click Show Welcome Guide… on the window's Home page to see the guide again.

1. Grant Microphone access so the app can record while the hotkey is held.
2. Grant Accessibility access so it can observe the push-to-talk hotkey and paste the result at the cursor.
3. Wait until the menu says Hold ⌃⌥ to dictate.

The first launch downloads and prepares these models:

• Parakeet speech recognition: about 600 MB
• Silero voice activity detection, which finds where you're speaking: about 1 MB, fetched after Parakeet
• Vedu Scribe transcript cleanup: about 560 MB

These downloads come from Hugging Face. They contain model files only; no audio or transcript is uploaded. Later launches use the cached copies.

You can dictate while Vedu Scribe downloads: HeyVedu pastes the raw transcript meanwhile and says so once. The menu shows the download's progress, and a notice appears when cleanup starts working.

Because source builds are ad-hoc signed, macOS may forget the Accessibility grant after rebuilding. See Troubleshooting if the hotkey stops working.

Use it

1. Put the text cursor where the dictation should appear.
2. Hold Control + Option together. Wait for the floating indicator to say Listening before speaking, especially with Bluetooth microphones.
3. Speak, then release both keys.
4. HeyVedu transcribes, optionally cleans the text, and pastes it at the cursor.

Press Escape while recording to cancel. A quick tap shorter than about 0.2 seconds, another key pressed with the chord, or an added modifier is treated as a keyboard shortcut and discarded. New dictations are ignored while the previous one is being processed.

Hands-free dictation

For longer dictation, tap the shortcut once instead of holding it. The floating indicator shows a lock while HeyVedu keeps listening. Press the shortcut once more to finish, or press Escape to cancel. You can type while it listens; only Escape cancels.

A hands-free recording also finishes by itself once you've been quiet for 10 seconds, after a short countdown in the indicator. Change that, or turn it off, with Stop when you go quiet in the window's Shortcut section. Either way it stops after 10 minutes. Turn off Tap for Hands-Free in the menu, or in the Shortcut section, to disable hands-free dictation; a quick tap then does nothing. While you hold the shortcut, the indicator reminds you that a tap starts hands-free.

Paste the last dictation again

If a dictation didn't land where you wanted, put the cursor in the right place and choose Paste Last Dictation from the menu bar icon, or Copy Last Dictation to put it on the clipboard. HeyVedu keeps only the most recent result, in memory, and forgets it five minutes after you dictated it.

When nothing that accepts text has focus (a button or a file list, for example), HeyVedu doesn't paste. It leaves the text on the clipboard and says so; press Command + V where you want it.

The text goes to the app you were in when you finished speaking. If another app comes to the front while HeyVedu is still transcribing, HeyVedu switches back before pasting. If that app has quit, the text is left on the clipboard instead.

Settings

Settings live in the HeyVedu window. Click the Dock icon, or choose Open HeyVedu… from the menu-bar icon. Its sidebar has Home (status, permissions and the welcome guide), General, Shortcut, Dictionary and Models. The menu-bar icon keeps quick toggles for cleanup, hands-free mode and the microphone.

Dictation shortcut

The default is ⌃⌥. To change it, open the Shortcut section of the window, click Change…, hold the modifier keys you want, and let go. Shortcuts use modifier keys only. A single key must be fn or a right-hand ⌘, ⌥ or ⌃, because the left-hand keys start too many other shortcuts.

A shortcut recorded with any right-hand key only responds to the exact keys you held. For example, Right ⌥⌘ ignores the left ⌥ and ⌘ keys. A shortcut recorded with left-hand keys only works with either side's keys.

macOS acts on the fn (🌐) key itself. If you choose fn, set System Settings → Keyboard → Press 🌐 key to to Do Nothing. The Home and Shortcut sections link there until you do.

Cleanup engines

Engine	Where it runs	Notes
Vedu Scribe	On your Mac	Default. HeyVedu's own cleanup model, a Qwen3.5-0.8B fine-tune (model card). English-only; downloads about 560 MB on first use.
Apple Intelligence	On your Mac	Uses Apple's on-device Foundation Model. Falls back to the raw transcript if the model is unavailable.

Choose the engine in the window's General section. The Models section shows each model's download and lets you start or retry it. Turn off Clean Up Transcripts to paste the raw, locally generated transcript. If a cleanup engine fails or rejects a result, HeyVedu pastes the raw transcript instead and tells you why.

Dictionary

Open the window's Dictionary section to list words HeyVedu should always write your way: names, product terms and jargon. For each word you can add other ways the transcript writes it under Also heard as, one at a time (press Return after each), such as “hey vedu” or “Hey, we do” for HeyVedu, or “teh” for the. Because each one is entered separately, an alternative can contain a comma. Punctuation between words doesn't matter when matching, so “hey we do” also catches “Hey, we do.”

After cleanup, every match is replaced with your spelling. If a cleanup engine changes a listed word, HeyVedu pastes the raw transcript for that part instead. The Try it field shows the result on a sample sentence.

The dictionary is saved in ~/Library/Application Support/HeyVedu/Dictionary.json and only holds what you type into it.

Snippets

Open the window's Snippets section to save text you insert often, such as a scheduling link, an address, or an email signature. Under When I say, type a trigger phrase; under Insert, the text (press Option + Return for a line break). Say the trigger anywhere in a dictation (“here's my calendly link”) and HeyVedu inserts the text exactly as you typed it, line breaks included. A dictation that is only the trigger becomes only the text. Matching works like the dictionary: any capitalization, whole words, punctuation between words ignored. Pick phrases you wouldn't say otherwise.

Snippets are saved in ~/Library/Application Support/HeyVedu/Snippets.json and only hold what you type into them.

History

History is off by default. Turn on Keep dictation history in the window's History section to keep a list of your dictations, newest first, with the date and time of each. Copy or delete any of them, or clear them all. Turning history off deletes everything saved.

History is saved only on this Mac, in ~/Library/Application Support/HeyVedu/History.json (readable by your user only), and keeps the latest 1,000 dictations. Besides the pasted text, each entry keeps the raw transcript and the cleanup model's output, so a dictation can be included if you choose to send feedback or report an issue. Nothing is sent anywhere automatically.

Microphone

Choose System Default or a specific input device from the Microphone menu, or in the window's General section. If a selected device is missing, HeyVedu falls back to the system default and shows a notice. If a device disappears while recording, it processes the audio captured so far.

Recording indicator and sounds

In the window's General section, Recording indicator puts the floating indicator at the bottom of the screen (the default), the top, or hanging from the camera notch. Notch uses the top of the screen on displays without one. Turn on Play sounds for a short sound when recording starts, finishes, or is cancelled.

During hands-free dictation the indicator shows how long you've been recording, and counts down for the last three seconds before stopping on silence. Speak to keep going.

Privacy

The fully local path is the default:

Microphone → in-memory audio → on-device Parakeet → on-device Vedu Scribe → paste

Data	Vedu Scribe / cleanup off	Apple Intelligence
Recorded audio	Memory only; stays on-device	Memory only; stays on-device
Speech recognition	On-device	On-device
Transcript cleanup	On-device	On-device
Dictation history	Off by default: the last result stays in memory for 5 minutes. Opt in to keep it on disk	Off by default: the last result stays in memory for 5 minutes. Opt in to keep it on disk
Personal dictionary and snippets	What you type only, on disk	What you type only, on disk

Text is inserted by briefly placing it on the macOS clipboard, issuing Command + V, and restoring the previous clipboard contents after roughly half a second. Clipboard managers are asked to treat this entry as transient, but their behavior is outside the app's control. When nothing that accepts text has focus, the text is left on the clipboard as a normal copy instead.

Debug builds write timing and error diagnostics to ~/Library/Logs/HeyVedu/debug.log. Audio, transcripts, and typed keys are not written to this log.

Troubleshooting

The menu says “Permissions needed”

Open System Settings → Privacy & Security, enable HeyVedu under both Microphone and Accessibility, then relaunch the app.

The hotkey stopped working after a rebuild

Ad-hoc signatures can cause macOS to retain a stale Accessibility entry. Quit HeyVedu, run:

tccutil reset Accessibility com.heyvedu.app

Relaunch the app and grant Accessibility access again.

A model failed to load

Check the internet connection and available disk space, then click Retry Loading Speech Model when that option appears. For Vedu Scribe, the next dictation retries its preparation. First-time Core ML compilation can take longer than later launches.

Dictation is transcribed but not inserted

Confirm Accessibility access is enabled. The app simulates Command + V, so the target application must accept paste at the current cursor position. The text isn't lost: choose Paste Last Dictation or Copy Last Dictation from the menu bar icon within five minutes.

Apple Intelligence is unavailable

Check that the Mac runs macOS 26 or later and supports Apple Intelligence, that it is enabled in System Settings, and that the system model has finished downloading. You can continue with Vedu Scribe or raw transcripts meanwhile.

Bluetooth audio starts late

Keep holding the hotkey until the HUD says Listening, then begin speaking. This gives the device time to switch into its microphone profile.

How it works

HeyVedu records 16 kHz mono audio in memory with AVAudioEngine. While you speak, Silero voice activity detection scores each quarter second for speech; silence before and after you speak is trimmed before recognition. FluidAudio runs Parakeet through Core ML for speech recognition. The selected cleanup engine normalizes the transcript, personal dictionary replacements are applied, then the app temporarily uses the clipboard to paste the result into the frontmost application.

Dependencies and model revisions are pinned in the project. Vedu Scribe files are also checked against pinned SHA-256 hashes before loading. When an app update pins a newer model revision, the app downloads it on next use and deletes the old one.

Current limitations

• English dictation only
• Apple silicon only
• Signed releases require a Developer ID certificate and notarization credentials
• Shortcuts are modifier keys only; a letter or Space can't be part of one
• Text insertion relies on Command + V and may not work in fields that block paste
• Restoring the clipboard may not preserve lazily provided clipboard content from every application

License

HeyVedu is licensed under the GNU General Public License v3.0. Third-party packages and downloaded models retain their own licenses. Vedu Scribe is released under Apache-2.0.