Private, local-first NVIDIA dictation for Linux.
VoicePad continuously persists microphone audio and transcribes it locally with NVIDIA's official Parakeet TDT 0.6B v3 model through PyTorch FP16 CUDA. Audio, transcripts, and model artifacts remain on the machine.
- Linux x86_64
- NVIDIA CUDA GPU in the tested 4 GB physical class or larger
- Python 3.13+
- A microphone available through the shared Linux sound server
There is no CPU ASR fallback and no silent precision or model fallback.
mise install
uv sync --upgrade
uv run voicepad prepare
uv run voicepadPress Space to record and Space again to stop. The verified model remains resident between recordings.
CLI alternatives:
uv run voicepad record start --duration 10
uv run voicepad transcribe path/to/recording.wavPackage publication is currently frozen. See
docs/content/designs/transcription-pipeline.md
for the complete architecture and validation evidence.
uv run ruff check
uv run ruff format --check
uv run ty check
uv run pytest packages --cov=voicepad --cov=voicepad_core --cov-fail-under=70
pnpm --dir docs install --frozen-lockfile
pnpm --dir docs build