Private by architecture
Microphone audio, model artifacts, and transcripts remain on your machine. No account, API key, or remote inference endpoint.
Local-first NVIDIA dictation
Speak naturally.
Keep it private.
VoicePad turns speech into text on your own NVIDIA GPU. Durable audio, resident inference, and a global shortcut—without sending a word to the cloud.
Linux x86_64 NVIDIA CUDA No cloud account
VoicePad is listening · 00:18Press the global shortcut to stop
Parakeet v3NVIDIA CUDA · FP16
recordhistorysettings
transcription
Dictation should feel immediate, private, and dependable. Your audio stays here while the model turns every phrase into text.
141.8ssample recording
~6.3sfinite pipeline
4 GBtested GPU class
100%local processing
Designed for dependable dictation
Fast is useful.
Trustworthy is essential.
VoicePad keeps capture, inference, and presentation separate so a slow model—or a failed job—never gets to decide whether your recording survives.
Audio first, always
Canonical WAV audio is written continuously. Existing recordings are never overwritten, modified, or deleted automatically.
One shortcut away
Let your Wayland compositor own the global key. VoicePad receives a secure local toggle and reports state through native notifications.
Resident pipeline
Ready before you are.
Parakeet loads and warms once, then stays resident while the TUI is open. CPU-side VAD finds natural boundaries; CUDA handles transcription.
Explore the architecture →01 · capturePersist canonical WAV
continuousDisk-backed from the first sample.
02 · planFind semantic pauses
boundedSilero VAD on lightweight CPU frames.
03 · inferTranscribe on CUDA
residentOfficial Parakeet FP16 safetensors.
04 · assemblePublish authoritative text
completeConservative overlap reconciliation.
Current production target
Built for Linux.
Measured on NVIDIA.
VoicePad currently prioritizes one reliable deployment over a matrix of silent fallbacks. Unsupported hardware fails clearly before recording starts.
Operating systemLinux x86_64
GPUNVIDIA CUDA · 4 GB class+
RuntimePython 3.13 · PyTorch FP16
ModelParakeet TDT 0.6B v3
Alpha, open for review
Make your next thought local.
Prepare the verified deployment, keep VoicePad resident, and dictate from anywhere on your desktop.