Cursor aware STT (Speech-to-Text)

No offline STT solution available online worked properly on my PC, so I built my own. This is a lightweight, offline, and secure Speech-to-Text solution optimized for Linux. This tool allows you to transcribe speech locally and inject it directly into any active application.

Key Features

• Toggle Trigger: Press and hold F8 to record, release to transcribe.
• 100% Offline: Uses whisper.cpp for local inference (no data leaves your machine).
• Inline Feedback: Displays recording... at your cursor while recording.
• Fast Injection: Uses ydotool for virtual keyboard input.
• Lightweight: Minimal background overhead using a low-level event daemon.

Prerequisites

This project is compatible with most Linux distributions (Fedora, Ubuntu, Debian, Arch, etc.) using systemd.

The installer will attempt to automatically install dependencies for:

• Fedora (dnf)
• Ubuntu/Debian (apt)
• Arch (pacman)

If you are on a different distribution, ensure you have these installed: alsa-utils, wl-clipboard, evtest, cmake, gcc/g++, make, libnotify, ydotool, git, curl.

Installation

1. Clone the repository:

   git clone https://github.com/AniruthKarthik/stt ~/stt
   cd ~/stt

2. Run the installer:

   make install

   Note: The installer will automatically:

   • Detect your distribution and install missing dependencies.
   • Clone and build whisper.cpp.
   • Download the base English model.
   • Configure background services (ydotoold and stt-daemon).
   • Set up necessary permissions and udev rules.

3. Finalize Permissions: If the installer added you to the input group, you must log out and log back in for the changes to take effect.

Usage

1. Start the Service: The installation script automatically enables the user service. If it's not running:

   systemctl --user enable --now stt-daemon.service

2. Record & Transcribe:

   • Open any text application (Terminal, Browser, Editor).
   • Press and hold F8.
   • You will see recording... appear at your cursor.
   • Speak clearly.
   • Release F8. The status text will be replaced by your transcription.

Architecture & Configuration

• Daemon: bin/stt-daemon listens to /dev/input/event* for F8 events.
• Logic: bin/whisperstt handles the recording (arecord), transcription (whisper.cpp), and typing (ydotool).
• Trigger Key: F8 was chosen as the default to avoid conflicts with modern "Copilot" keys, which often send complex macros (Meta+Shift+F23) that can interfere with synthetic typing.

Troubleshooting

• No Typing: Ensure ydotoold is running (systemctl status ydotoold.service).
• Permission Denied: Verify your user is in the input group (groups).
• Logs: Check detailed output here:

  tail -f ~/stt/tmp/stt.log