Whispy (Voice Dictation for macOS & Linux)
· whispy-dun.vercel.app
Whispy is a local voice dictation utility for macOS and Linux (X11). It runs
NVIDIA's Parakeet speech model on your own machine, so no audio or text ever
leaves your computer. Hold a trigger key (Fn on macOS, Right Ctrl on
Linux) to record and release it to type the transcription into the active
field, or switch on toggle mode to press once to start and once to stop.
Install
curl -fsSL https://raw.githubusercontent.com/albrones/whispy/main/scripts/bootstrap.sh | bash
On macOS this builds and installs the signed Whispy.app. On Linux, install the
system packages first (see the install guide below). If something does not
work afterwards, run make doctor from the repository for a diagnosis.
Documentation
The full guides live on the website:
Platform support
Whispy runs on macOS and Linux (X11). The OS-coupled seams sit behind a
ports-and-adapters layer (src/whispy/platform/) bound at runtime by
platform.detect().
Not supported in v1: Wayland (global hotkey and synthetic input are
restricted by its security model; pick an "Xorg"/"X11" session at login),
Windows, native Linux packaging, and the floating overlay window on Linux
(state is shown through the tray instead).
The model
Whispy ships one model: nvidia/parakeet-tdt-0.6b-v3, int8-quantized to
ONNX and run on CPU through onnx-asr.
There is nothing to choose, so there is no model setting.
- Download: 639 MB, automatic on first use
- Languages: 25, detected automatically per utterance. Speech outside this
set will not transcribe: Bulgarian, Croatian, Czech, Danish, Dutch, English,
Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian,
Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian,
Spanish, Swedish, Russian, Ukrainian (see the
model card).
Contributing
Contributions and design input are welcome. Development setup, the test suite
and how to regenerate the docs are in CONTRIBUTING.md.
License
This project is distributed under the GPLv3 license. See the LICENSE file
for the full text.
The transcription model is a separate work under a different licence:
nvidia/parakeet-tdt-0.6b-v3 is © NVIDIA Corporation, licensed
CC-BY-4.0. Whispy does not
redistribute it: the weights are downloaded from the Hugging Face Hub on first
run and cached on your machine, so no model files ship in this repository or in
any release artifact. See NOTICE for the full attribution, including the ONNX
conversion Whispy loads.
