Voice Dictation

• LilySpeech, works better than Windows Speech Recognition (buggy)

  • Voice Commands: https://lilyspeech.com/dictation-voice-commands/
  • if stops working, restart the app, after killing it if needed
  • Add/modify Custom Word Replacements
    • c:\Users\Admin\AppData\Local\LilySpeechApp\LilySpeechUser\replacements.txt

• WhisperFlow

  • $12/mn but uses AI to automatically format punctuation and more

• Whisper.CPP Server - NOT an effective solution

  • delay was too slow
  • runs locally (offline), for any Windows or WSL app that accepts keyboard input
  • Hotkey: Alt+Space — press once to start recording, press again to stop; transcription is pasted at cursor
  • Installation
    • Script location: D:\FSS\Software\Utils\whisper-hotkey\whisper-hotkey.py
    • Silent launcher: D:\FSS\Software\Utils\whisper-hotkey\whisper-hotkey-launch.vbs
    • Whisper server (runs in WSL, auto-starts): installed via uvx voice-mode — reinstall with uvx voice-mode whisper install — no backup needed
    • Windows Python dependencies (one-time): cd D:\FSS\Software\Utils\whisper-hotkey then uv sync — installs all dependencies into .venv
    • Microphone privacy check: Settings → Privacy → Microphone → ensure “Allow apps to access your microphone” is ON. After changing, restart WSL: run wsl --shutdown in PowerShell.
    • Auto-start at Windows login: A shortcut already exists in the Startup folder (whisper-hotkey.lnk) pointing to the launcher — nothing to copy. Verify: Win+R → shell:startup
  • Usage
    • Press Alt+Space to start recording (script must be running)
    • Speak naturally — punctuation is added automatically
    • Press Alt+Space again to stop — text is pasted at cursor via Ctrl+V
    • Works in: Windows Terminal, Obsidian, browsers, Word, Claude Code — any app accepting keyboard input
    • There is a short delay after stopping before text appears (transcription time)
  • Notes
    • Output uses clipboard paste (Ctrl+V), not keystroke injection — this ensures compatibility with Electron apps (Obsidian, VS Code)
    • The Whisper server runs in WSL at localhost:2022. If it’s not reachable, open WSL and run uvx voice-mode whisper start
    • The base Whisper model is used by default (fast, ~223MB RAM). For better accuracy at the cost of speed, change to small or medium in ~/.voicemode/voicemode.env → VOICEMODE_WHISPER_MODEL=small
    • Also used by VoiceMode MCP in Claude Code for two-way voice conversations (same local server)