yazses

yazses [options] [command] [args]

yazses is an offline voice dictation and speech-to-text tool for Linux (X11 and Wayland), macOS, and Windows. You hold a hotkey, speak, and release; on-device faster-whisper transcribes the clip and the text is typed into the focused window. There is no account, API key, or network call in the default path.It also transcribes existing recordings (yazses transcribe) and can capture a whole meeting with speaker labels (yazses meeting). Dictation works out of the box. Speaker diarization and local-LLM meeting minutes are extras, off by default.On Linux, yazses setup installs PortAudio, injection tools (xdotool on X11, ydotool / wtype on Wayland), and adds the user to the input group so the hotkey can be read from the kernel. Joining that group only takes effect after a full log-out and log-in.The daemon loads the speech model once (first start can take 10–30 seconds) and then listens for the hotkey. yazses start prefers a systemd user unit when one is installed; otherwise it detaches. yazses-daemon runs the same process in the foreground for debugging.

-h, --help

Show grouped command help. Every subcommand also accepts -h / --help.

Print the installed version.

Install shell tab completion for commands and options.

Print the completion script instead of installing it.

Print the next three steps for this machine. Changes nothing.

Verify platform, microphone, keyboard capture, injection backend, model cache, and config. Each line is OK / WARN / FAIL / SKIP.

On Linux, install audio and injection packages, join the input group, and enable ydotoold on Wayland. --dry-run previews the changes.

Control the dictation daemon. start restarts an already-running instance instead of spawning a duplicate.

Show state, hotkey, model, injection backend, uptime, and recent decode latency. --json for scripts.

Show or change the hold-to-talk key (right_alt, right_ctrl, space, …). Default auto is right_alt on Linux. Restart the daemon after changing it.

Transcribe a local audio or video file offline. --format / -f is txt (default), md, srt, vtt, or json. --out / -o sets the output path. --model, --language, and --diarize override the dictation config for this run.

Hands-free meeting capture. Off by default; enable with yazses features enable meeting.

List capabilities and turn them on or off without editing the config file.

Install or remove a systemd user unit that starts YazSes at login.

Record short utterances and write microphone / VAD thresholds to the config.

Print the diagnostic log (metadata only, not dictated text). -n sets the line count.

Settings live in ~/.config/yazses/config.toml on Linux. Every section has defaults, so a missing file is valid. Prefer yazses features enable / disable over hand-edits; that path writes the right keys and keeps comments.

$ [stt]
engine = "faster-whisper"
model = "base.en"
language = "en"
device = "cpu"
compute_type = "int8"

[hotkey]
key = "auto"
hold_threshold_ms = 500

[audio]
device = ""
max_record_seconds = 300

[injection]
backend = "auto"

First use downloads a speech model (on the order of a hundred megabytes for base.en). Linux hotkeys stay dead until you log out after yazses setup joins the input group; a new terminal tab is not enough. Meeting capture and --diarize need extra packages and models. Default checkpoints are English-first. The tray icon needs a system tray host (on GNOME, the AppIndicator extension). Experimental features refuse to enable unless you pass --force.

YazSes is an Apache-2.0 project by Mohsen Seyedkazemi Ardebili. It is built on faster-whisper (CTranslate2) for on-device transcription and aims at hold-to-talk dictation without a cloud account.

whisper(1), faster-whisper(1), ffmpeg(1), ydotool(1), xdotool(1), wtype(1)

Source code · Documentation