Local production processing
The customer build contains no cloud processor, account, advertising, analytics, or tracking.
Local voice typing
TalkDo turns speech into clean local dictation in the Mac application you are already using.
If the destination changes while TalkDo is working, the text is held safely in the menu instead of being pasted into the wrong place.
Clean Dictation removes genuine fillers and accidental repetition, fixes punctuation and sentence boundaries, and preserves your wording and meaning. It uses deterministic local rules, never a generative rewrite.
| Action | Shortcut | Behavior |
|---|---|---|
| Dictate | Option Space | Hold, speak, release |
| Paste last transcript | Option Shift V | Insert the latest result again |
The dictation and paste-last shortcuts can be changed in Settings. Dictation always uses hold, speak, release.
App-aware protection recognizes terminals, coding tools, and editors. It preserves identifiers such as camelCase, snake_case, command flags, environment variables, paths, and filenames without changing your selected output.
WhisperKit downloads the recommended model once and then runs offline. TalkDo uses macOS System Default for microphone routing.
The customer build contains no cloud processor, account, advertising, analytics, or tracking.
Choose how long transcripts and failed recordings remain, or delete all TalkDo data from Settings.
No Prompt, cloud processing, Command Mode, clipboard monitoring, Capture inbox, or hidden background modules.
The compact click-through indicator shows input level, trailing words, destination, and each processing stage, then disappears when work completes. Choose any top or bottom Left, Center, or Right anchor directly from the menu or in General; the position persists and updates live.
TalkDo follows macOS System Default instead of rebinding Core Audio devices. A zero-frame health check stops dead capture rather than leaving Listening active.
TalkDo starts at login and reopens after an unexpected stop. A normal Quit stays quit, and rapid failures stop instead of creating a crash loop.
The Hub and menu say Ready only when dictation can insert into other apps. Missing Microphone or Speech permission is an action item; missing Accessibility is clearly clipboard-only.
Run Setup Check from Settings. It passes only after fresh hotkey dictation reaches TalkDo's practice field; typing cannot imitate success. Only the result and time are retained.
Every run uses the same local cleanup path without model selection, generative rewriting, authentication, network, or token cost.
Changed History rows open a local side-by-side review of raw speech recognition and final output, with a bounded word-edit summary and separate copy actions. Raw recognition is clearly not labeled as human-verified audio.
Search intended spellings or heard-as variants in Words, then star important names and technical terms. Starred entries lead the bounded local recognition vocabulary; TalkDo does not auto-learn ordinary text.
On prose surfaces, TalkDo uses the immediate cursor boundary to add only missing outer spaces. Terminal, code, paths, URLs, flags, and identifiers remain exact.
A downloaded local model prepares quietly after launch. The Hub reports preparation honestly, and the cache-only startup path never initiates a download.