Show HN: GHOSTYPE – AI voice input that learns your writing style
1 コメント
The "Virtual Personality Engine" / style transfer is a killer feature. One of the biggest issues with LLM-generated content (whether for voice, text, or even brand answers) is the "generic AI accent" that just screams "I was generated by an LLM".
Curious how you handle the "style vector" update frequency? Does it continuously learn as I dictate more, or is it a static snapshot? For brand/personal voice consistency, continuous learning would be huge but might drift.
Also, +1 for the retro CRT aesthetic.
Curious how you handle the "style vector" update frequency? Does it continuously learn as I dictate more, or is it a static snapshot? For brand/personal voice consistency, continuous learning would be huge but might drift.
Also, +1 for the retro CRT aesthetic.
I’m the creator of GHOSTYPE.
I built this because I wanted a voice input workflow that didn't feel like talking to a chatbot. I wanted something that felt like a "neural extension" of my keyboard.
It’s a macOS-native app that sits between your voice and your active window. Here is what it actually does:
The Core Input Workflow:
Push-to-Talk & Smart Send: Hold a global shortcut to speak. It detects the active app to determine the correct send method (e.g., Cmd+Enter for Slack vs Enter for Discord).
Inline Editing: You can format while speaking. Need a line break, a specific spelling, or a bulleted list? You just say it, and it handles the formatting inside the sentence before outputting.
"Call Ghost": Post-processing commands (translate, polish, expand) are available immediately after speaking, before the text is typed out.
The Experimental Stuff (WIP):
1. Ghost Twin (Style Transfer): I call this a "Virtual Personality Engine" (a bit pretentious, I know). It analyzes your local writing history to build a style vector. It learns your tone—whether professional for emails or casual for Discord—so the output sounds like you, not a generic LLM.
Side note: I'm currently building the training UI to look like a retro CRT terminal because I miss that aesthetic.
2. Ghost Morph (Custom Skills): Trigger custom macros with a modifier key. For example, turn a raw voice thought directly into a formatted Twitter thread or a Jira ticket structure.
Privacy & Architecture:
Local-First: I’m an indie dev and literally cannot afford the server costs to store user typing data.
E2E Encrypted: Your "Ghost Twin" profile and skills are synced across devices via E2E encryption. Data stays on your machine.
Constraints:
macOS Only: I don't own a Windows PC, so I can't build for it yet. (If any Windows devs want to port this, let me know).
It's currently in pre-launch. I threw up a landing page at if you want to follow the progress. https://ghostype.one
Happy to answer questions about the local-first architecture or the style transfer logic!