Changelog

Notable changes to Herga for users. Each release gets a section here, newest first. The website shows this file at herga.mrgnhnt.com/changelog.

Unreleased

The first release of Herga as a dictation app. This fork of jamiepine/voicebox 0.5.0 drops text-to-speech and focuses entirely on turning your speech into ready-to-send text on Apple Silicon Macs.

Voicebox is now Herga, from jerga, Spanish for slang. Your captures, styles, dictionary and settings carry over the first time Herga opens. macOS treats it as a new app, so it asks for Microphone, Accessibility and Input Monitoring once more. The old Voicebox.app can go in the Trash; the install script does that for you.

New

Guided setup. First-run onboarding downloads the models in the background while it walks you through permissions, your hotkey, your name and a first messy dictation. If macOS needs Herga to quit for a permission, setup reopens on the same step.
Native, streaming dictation. The microphone is captured in native code and streamed while you speak, and cleanup runs a sentence at a time, so there’s less to wait for when you let go.
Text lands where you were typing. Text is inserted through Accessibility, typed keystrokes or the clipboard, whichever works for the app. When it has to use the clipboard, it puts yours back.
Writing styles per app. Every app gets a style, and each style learns on its own. Switch style by saying so at the start of a dictation (“use formal mode”). The pill shows the style’s name and plays a sound when it changes.
Teach by replying. Teach a style how you write by replying to a few short conversations.
Dictionary. Add terms and spoken replacements for everywhere, a style or specific apps. Select a word in a capture to add it.
Command Mode. Select text in any app, hold right ⌘ + right ⇧ and say how to rewrite it. It comes with the Polish and Prompt Engineer transforms, and you can save your own.
Spoken commands. Say line breaks, lists, quotes, brackets, braces, slashes, pipes and carets. Spell things out letter by letter (“capital C…”). Say “paste from clipboard” to insert your clipboard.
Self-corrections. “Tuesday, no actually Wednesday” becomes “Wednesday”. Repeats, stutters and restarts are removed.
Correction learning. Fix a capture and the fix is used right away. Older fixes are folded into rules, which are kept only if they do at least as well on your past corrections.
Captures redesign. Captures are grouped by app and show the audio next to what Whisper heard and the cleaned-up text. A Check badge flags cleanups worth a second look.
Insights. See words dictated, speaking pace, time saved and your most-used apps.
History retention. Choose how long to keep captures. Nothing is deleted until you confirm.
Escape to cancel a dictation, even while it’s still finishing.
Sound cues for start, stop and errors, with a volume setting.
Launch at login, on by default, with the window hidden.
Command palette. Press ⌘ K to jump to any setting or action.

Improved

After an update, Herga replaces a local server still running from the old version instead of reusing it, so the new version’s backend is always the one running.
Model weights stay loaded between dictations, and the first dictation after launch is as fast as the rest.
The pill shows on the display you’re working on, shows recording as soon as you press the keys, and never takes keyboard focus.
Audio without a voice in it is never transcribed, and Whisper’s repeated-phrase loops are removed.
Cleanup is rejected and your own words are used instead when it answers your question instead of writing it down, copies an example, or adds words you didn’t say.
The microphone is released after every dictation.

Removed

Text-to-speech, voice profiles and cloning, stories, effects, the MCP server, and Herga Cloud.
Support for Windows, Linux and Intel Macs. Herga now runs only on Apple Silicon.