hush/README.md

Hush

A Linux speech-to-text daemon that records audio, transcribes it via the Groq Whisper API, optionally cleans up the text with an LLM, and types the result into whatever window currently has focus.

The intended workflow: bind hush toggle to a global hotkey, press to start recording, press again to stop, and the transcribed text appears in your active application.

Requirements

Runtime dependencies:

• pw-record (PipeWire) or ffmpeg (PulseAudio) — audio capture
• wtype (Wayland) or xdotool (X11) — text injection
• notify-send — desktop notifications
• A Groq API key

Build dependencies:

• .NET 10.0 SDK

Installation

sudo ./install.sh

This builds the project, installs the hush binary to /usr/bin/hush, sets up zsh tab-completion, and registers hush.service as a systemd (or OpenRC) user service.

To uninstall:

sudo ./uninstall.sh

Setup

After installing, run the interactive setup wizard to configure your API key and preferences:

hush setup

Config is stored at ~/.config/hush/config in TOML format.

Commands

Command	Description
hush toggle	Start recording if idle, stop and transcribe if recording
hush start	Start recording
hush stop	Stop recording and transcribe
hush abort	Discard the current recording
hush status	Show daemon state and recording duration
hush show	Show current configuration
hush latency-test	Measure STT and LLM round-trip latency
hush daemon	Run the daemon in the foreground
hush setup	Interactive configuration wizard

Building Manually

./build.sh

Produces a native AOT binary at Hush.Cli/bin/Release/net10.0/linux-x64/publish/Hush.Cli.

Configuration

Key settings in ~/.config/hush/config:

Key	Default	Description
groq_api_key		Groq API key
whisper_model	whisper-large-v3-turbo	Whisper model to use
llm_model	openai/gpt-oss-20b	LLM model for text cleanup
audio_backend	pw-record	pw-record or ffmpeg
typing_backend	wtype	wtype or xdotool
whisper_language		Optional ISO 639-1 language code
min_recording_duration	500	Minimum recording length in ms