Zero dependency, pure C++ speech-to-text binary for Linux, done the UNIX way (daemonless, no bloat, no slop, no GUIs, no venv, nothing)

This is just a very simple, 100% local STT toggle/CLI tool (open source & Apache-2 licensed) that adheres to the UNIX philosophy, does one job and one job only.

Tap once, speak for as long as you want, tap again, transcribed and copied to the clipboard.

A native C++ binary that links the whisper.cpp C API directly (pulled from a pinned commit, GGML models are downloaded from Hugging Face).

Everything else you already have.

No deps beyond standard C++ and Linux. If you have a C++ build environment on Linux you almost certainly have everything you need already.

Also, it's CPU only.

CUDA? Vulkan? GPU backend? The baseline question is, does this 3D object contain an ancient artifact known as a CPU? If yes? Then it will work.

The binary is a stateful toggle, with a very simply and tiny CLI surface:

asryx # Toggle record/transcribe asryx status # Check idle/recording/transcribing asryx --language <auto|CODE> # Set language asryx --model list # List supported models asryx --model install <MODEL> # Download model asryx --model use <MODEL> # Switch model

Default model is base.en at 142 MiB.

But works with all supported GGML langs, which cover a 100 languages.

And since it's a toggle you can keybind it, for example on Hyprland I have it like this:

bind = ALT, W, exec, asryx

You can hook it up to Sway, i3, GNOME, etc.

The way it works TL;DR:

First keypress captures audio via PipeWire or ALSA.

Second keypress stops capture, runs inference in-process, copies to clipboard, wipes temp files, exits.

Doesn't stay in memory between uses.

Doesn't load the model unless invoked.

Boots instantly & exits instantly.

One command to install (YOU compile it on YOUR own machine, no pip install questionable-library, or cargo install questionable-crate).

One command uninstall + the README lists every file and folder the tool touches.

It removes all runtime artifacts before exiting. The idle footprint is exactly 0MB.

And it basically never errors out as long as your machine has a light source.

There is no daemon, no server, no queue, no background service, and no moving state outside the current toggle.

Every run goes through one lock directory and live PID checks first, so double taps, compositor repeat, or accidentally hitting the key 10 times collapse into safe no-ops instead of spawning 10 recorders.

Source --->