Dependencies Reference

Runtime dependencies

Package	Version	Purpose
faster-whisper	>=1.2, <2.0	Speech-to-text via CTranslate2 (Whisper engine)
piper-tts	>=1.4, <2.0	Local neural text-to-speech
sounddevice	>=0.5, <1.0	Audio I/O — microphone capture and speaker playback
numpy	>=1.26, <3.0	Numerical operations for audio processing

Optional dependencies

Package	Version	Purpose
scipy	>=1.11	High-quality polyphase resampling for Whisper downsampling. Falls back to boxcar filter if not installed.

Install with: pip install -e ".[quality]"

Development dependencies

Package	Version	Purpose
pytest	>=8.0	Test framework
ruff	>=0.6	Linting and formatting

Install with: pip install -e ".[dev]"

System dependencies

Package	Purpose	Required
PortAudio	Audio backend for sounddevice	Yes (pre-installed on most Linux desktops)
CUDA 12 drivers	GPU acceleration for Whisper	Yes for GPU mode
nvidia-cublas-cu12	CUDA 12 BLAS library (via pip)	Yes for GPU mode
nvidia-cudnn-cu12	CUDA 12 DNN library (via pip)	Yes for GPU mode
notify-send	Desktop notifications (libnotify)	Optional
xdg-open	Open URLs in default browser	Optional (for conversation viewer)

External dependencies (not in pyproject.toml)

Component	Purpose	Installation
OpenCode CLI	LLM backend	Installed by install.py or manually
Piper voice models	TTS voice data	Downloaded by install.py from Hugging Face
Whisper model weights	STT model data	Downloaded automatically by faster-whisper on first use

CUDA library path

CTranslate2 (used by faster-whisper) requires CUDA 12 runtime libraries. These are installed via pip inside the venv's site-packages, not in the system library path. The run_daemon.sh script adds them to LD_LIBRARY_PATH:

export LD_LIBRARY_PATH="$SITE/nvidia/cublas/lib:$SITE/nvidia/cudnn/lib:${LD_LIBRARY_PATH:-}"

This is why you must use run_daemon.sh (or set LD_LIBRARY_PATH in your service file) instead of running python -m mirach directly for GPU mode.