Daemon Whisper

"When you whisper into the daemon, the daemon types back"

CPU-only local speech-to-text transcription for Linux using Whisper models. Fully offline - your voice never leaves your machine.

Quick Summary

• Purpose: Convert voice to text locally on Linux
• Backend: whisper.cpp with CPU-only support
• Activation: Mod+V keybinding (configurable)
• Output: Auto-typed at cursor using wtype
• Model: ggml-small.en.bin (466MB) - can switch to ggml-base.en.bin for speed

How It Works

1. Press Mod+V → Start recording (notification appears)
2. Speak clearly into your microphone
3. Press Mod+V again → Stop and transcribe
4. Text appears at your cursor automatically

Features

• Toggle recording: Same key starts/stops recording
• Visual feedback: libnotify notifications for recording status
• Auto-type integration: Uses wtype to type at cursor position
• Error handling: Gracefully handles no-speech detection
• CPU-only: No GPU required, works on any Linux system
• Fully offline: Your voice data never leaves your machine

CPU-Only Operation

This project is designed for CPU-only operation with no GPU acceleration. The build configuration explicitly disables Vulkan and GPU support:

cmake -DWHISPER_VULKAN=OFF ..

This ensures maximum compatibility and stability across all hardware configurations, including systems with Intel iGPUs or no dedicated GPU. Processing times listed below reflect CPU-only performance.

Performance

Size	Parameters	English-only	Multilingual	Required VRAM	Relative Speed
tiny	39 M	tiny.en	tiny	~1 GB	~10x
base	74 M	base.en	base	~1 GB	~7x
small	244 M	small.en	small	~2 GB	~4x
medium	769 M	medium.en	medium	~5 GB	~2x
large	1550 M	N/A	large	~10 GB	1x
turbo	809 M	N/A	turbo	~6 GB	~8x

Model Selection Notes:

• The .en models (English-only) perform better than multilingual models for English transcription
• The difference is most significant for tiny.en and base.en, less significant for small.en and medium.en
• The turbo model is an optimized version of large-v3 offering faster transcription with minimal accuracy degradation
• Processing times reflect CPU-only performance (slower than GPU benchmarks)

Recommended Models

• tiny.en (~78MB) - Fastest, basic accuracy
• base.en (~148MB) - Fast, good accuracy
• small.en (~466MB) - Balanced, very good accuracy [recommended for most users]
• medium.en (~1.5GB) - Slower, excellent accuracy
• large (~3GB) - Best accuracy, multilingual support
• turbo (~1.6GB) - Fast + excellent accuracy (optimized large-v3)

System Requirements

• Linux OS with PipeWire or PulseAudio
• Window Manager (Niri, Hyprland, i3, Sway, etc.) or Desktop Environment (KDE, GNOME)
• Dependencies:
  • bash - Shell interpreter
  • cmake + build tools - For building whisper.cpp
  • ffmpeg - Audio processing
  • alsa-utils - Provides arecord for audio recording
  • wtype - Auto-typing at cursor (Wayland)
  • libnotify - Desktop notifications
• ~500MB disk space - For whisper.cpp and small.en model

Limitations

• English-only: Transcription is hardcoded for English language
• Wayland-focused: Uses wtype for auto-typing (Wayland). For X11, replace with xdotool
• CPU-only: No GPU acceleration (by design for stability)

Installation

See INSTALL.md for detailed installation instructions.

Configuration

See KEYBINDINGS.md for keybinding configuration.

TUI Companion App (Optional)

An optional TUI companion app is available for model management and performance tracking. See TUI.md for details.

Features:

• View and download Whisper models
• Switch between installed models
• View transcription statistics and CPU usage
• System information display

Installation:

pip install rich psutil
cp whisper-tui.py ~/.local/bin/whisper-tui
chmod +x ~/.local/bin/whisper-tui

Run with: whisper-tui

License

MIT License - see LICENSE for details.

Contributing

Contributions are welcome! Feel free to submit issues or pull requests.