speakcode 0.1.0

Private voice-to-text for developers. Hold a key, speak, get text pasted into any app.

Voice Coding

Private voice-to-text for developers. Hold a hotkey, speak, get text pasted into any app.

• Private — audio goes to Gemini Flash (your own API key), nowhere else
• Fast — ~1 second transcription via Gemini 3.0 Flash
• Universal — auto-pastes into any focused app: VS Code, Terminal, Slack, browser, etc.
• Coding-aware — "dot env" → .env, "camel case foo bar" → fooBar

Install

pip install speakcode

Setup

1. Get a Gemini API key
2. Set your API key:

   export GEMINI_API_KEY=your_key_here
   

   Or create a ~/.voice-coding/.env file:

   GEMINI_API_KEY=your_key_here
   

macOS Permissions

Your terminal app (Terminal.app / iTerm / VS Code) needs two permissions in System Settings → Privacy & Security:

• Microphone — for audio recording
• Accessibility — for global hotkey detection and auto-paste keystroke simulation

After granting Accessibility, restart your terminal app for the permission to take effect.

Learn Project Vocabulary (Optional)

Teach Voice Coding the vocabulary of any project you work on:

cd /path/to/your/project
speak learn

This scans the repo (README, package.json, etc.) and merges its vocabulary into your global memory at ~/.voice-coding/memory.md. Run it in each repo you work on — terms accumulate across projects.

The memory file includes:

• Vocabulary — project-specific terms with disambiguation hints (e.g., "Claude Code" not "clock code")
• Context — brief descriptions of your projects and tech stacks
• Notes — space for personal customizations (accent, language mixing, corrections you've noticed)

Edit ~/.voice-coding/memory.md anytime to add or fix terms.

Usage

speak

Hold Alt (⌥) to start recording. Release Alt to stop, transcribe, and auto-paste into whichever app is focused.

Press Ctrl+C to quit.

Tips

• Speak naturally — filler words (um, uh, like, you know) are automatically removed
• Minor grammar is corrected while preserving your original wording
• Recordings shorter than 0.5 seconds are ignored to prevent accidental triggers

Coding Transforms

Voice Coding post-processes transcriptions with coding-aware rules:

You say	You get
"dot env"	.env
"slash api"	/api
"camel case foo bar"	fooBar
"snake case my variable"	my_variable
"open paren"	(
"arrow"	=>
"triple equals"	===
"new line"	newline character

How It Works

1. A macOS CGEventTap listens for the Alt key globally (works in any app, including VS Code)
2. sounddevice captures mic audio at 16kHz mono while the hotkey is held
3. Audio is sent to Gemini 3.0 Flash for transcription, with vocabulary from ~/.voice-coding/memory.md if present
4. Post-processor applies coding-aware text transforms
5. Result is copied to clipboard via pbcopy and pasted via osascript Cmd+V simulation