tornikegomareli/Talkify

Everything is on-device: Apple's SpeechAnalyzer/SpeechTranscriber for recognition, AVSpeechSynthesizer for Read Aloud, Translation for translation, and FoundationModels for prompt shaping. That last one is a language model, and it is Apple's, running here: there is no key, no account, and no request. Talkify makes no network requests, stores no audio, and keeps no history beyond the local usage metrics you can see in Insights.

Two caveats. Read Aloud reads a selection through Accessibility where it can, and by copying it where it cannot, which is any web page: the selection passes through the clipboard and the previous clipboard is put back afterwards.

And dictated text is inserted by pasting it, so it passes through the system clipboard for up to about half a second before your previous clipboard is put back. A clipboard manager or Universal Clipboard can see it during that window.

• macOS 26 (Tahoe) on Apple Silicon

With Homebrew:

brew install --cask tornikegomareli/talkify/talkify

The full name is what makes it one line. Homebrew 6 will not load anything from a third-party tap until you trust it, but installing by full name adds the tap and trusts this one cask on its own — nothing else in the tap, and no separate brew trust. Read the cask first if you would rather see what it does.

Or grab Talkify.dmg from the latest release.

Build from source:

git clone https://github.com/tornikegomareli/Talkify.git
cd Talkify
open Talkify.xcodeproj   # ⌘R

Or headless:

xcodebuild -project Talkify.xcodeproj -scheme Talkify -configuration Debug build

Run the tests the same way CI does:

xcodebuild test -project Talkify.xcodeproj -scheme Talkify -destination 'platform=macOS'

The trigger and the Read Aloud shortcut are rebindable in Settings → Shortcuts. Direct Dictation can use a keyboard key, Middle Click by pressing the scroll wheel, or another auxiliary mouse button. A bound button keeps its usual action unless the exact combination you bound is pressed, so binding ⌥ + Middle Click leaves plain middle click alone.

Drag an audio or video file to the top of the screen and the island opens to take it. It transcribes in the background, so fn keeps working, and the menu bar ghost shows progress.

When it finishes the island comes back holding the transcript. Drag it to a folder for the .txt, or to a text field for the words. Click to copy. Leave it and after five seconds it saves next to the source file, or into a folder you set in Settings → Drop Transcription. Hovering pauses that timer.

With a second dictation language configured, the target splits in two and the half you drop on picks the language. Transcribe File… in the menu does the same with a picker.

Pick a language in Settings → Language and the Translate key writes in it. Hold right ⌘, say it in English, and Spanish lands in the document. The key does not change when the language does, so there is one shortcut to remember.

Translation runs on your Mac through Apple's Translation framework, and the notch shows the pair before you speak. If it fails, nothing is pasted and your words go to the clipboard.

It works the other way too. Turn on Translate before speaking in Settings → Read Aloud, select text in a language you do not read, and the Read Aloud key speaks it in your voice's language. The voice is the target, so there is nothing else to set.

Turn on Shape dictation with a prompt in Settings → Prompt Shaping and the prompt you pick rewrites finished dictation through Apple's on-device model before it is inserted. Three prompts are built in, tighten grammar, bullet lists and remove filler words, and the library shows each one's instruction under its name so you pick by reading rather than by remembering.

Edit any of them, or write your own. The instruction is the whole prompt for most of them; a closing instruction and a worked example are there under Advanced when you want them. You can also try a prompt before you use it: type a sample sentence, press Try, and see what the model does with it. It runs the same on-device model a real session runs, so what you see is what you will get. The framing that keeps the model rewriting your words instead of answering them is fixed and not editable.

While you dictate, a caption under the notch names the prompt the session will shape with, colored from your Edge Glow palette and moving with your voice, and the bare arrow keys cycle through your prompts — or None, to insert the words exactly as spoken. After you release, the island stays up and the same caption names the prompt while the rewrite runs. Shaping is a beta and fails safe: any error, or an answer slower than ten seconds, inserts your raw words unchanged, history keeps what you actually said, and nothing leaves your Mac.

Pick a second language in Settings → Language and it gets its own trigger. Hold fn for English, hold right ⌥ for German, with no setting to change in between. Both triggers are rebindable, and either can use a keyboard key or a supported mouse button.

Code is organized into folders, callbacks only flow one way from the main wiring point, and a pure reducer handles the state.

• App/: composition root, settings store, status item
• Input/: the global key event tap and recorded bindings
• Dictation/: the session machine (DictationSessionMachine, a pure tested reducer), the speech/insertion services, and the HUD surface and voice visuals only dictation draws
• Translation/: the coordinator both callers talk to, the rules, and the two files that import Apple's Translation framework
• Prompt shaping lives in Dictation/: the prompt model, the service that races the on-device model against a timeout, and the caption the HUD draws
• DropTranscription/: the drag gesture, the file transcription service, where a transcript is staged and lands, and its own HUD surfaces
• CoreHUD/: what both features share: HUDStage owns the single panel and decides who holds the shape, plus the geometry seams and Metal shaders
• ReadAloud/, Settings/, Insights/

CONTEXT.md is the domain doc, decisions live in docs/adr/, and CONTRIBUTING.md is the contract for changing any of it

• Live Captions & Meeting Transcripts. Ephemeral captions from a selected app's audio (Chrome, YouTube, meeting apps), and the saved, timestamped transcript as a separate action. The domain design already lives in CONTEXT.md; the recognition pipeline is ready for non-microphone audio.
• Text cleanup. Shipping as the prompt shaping beta above — on-device rewriting of finished dictation, off by default. Per-application profiles and a benchmarked default remain future work; with shaping off, Talkify still inserts exactly what you said.
• Snippets. Saved text blocks inserted by a spoken trigger word: say your trigger mid-dictation and the whole block lands instead.

Read CONTRIBUTING.md first. It covers the issue-first flow, branch and pull request naming, code style, the AI policy, and what a bug report needs to contain to be actionable.

Start from an issue. A pull request that arrives without one behind it may be closed, however good the code is, because design belongs in the issue where it is still cheap to change.

MIT