Open source · Apache-2.0 · runs on your CPU
Hold a key, speak, release — YazSes transcribes your words on-device with faster-whisper and types them into whatever app has focus. No cloud. No API key. No subscription. Nothing leaves your machine.
bash <(curl -fsSL https://raw.githubusercontent.com/MSKazemi/yazses/main/install.sh) brew tap MSKazemi/yazses && brew trust MSKazemi/yazses && brew install --cask yazses winget install MSKazemi.YazSes pipx install yazses
base.en, CPU only, no GPU Quick start
Three steps, a few minutes, no account. The speech model downloads once; after that, dictation never touches the network.
bash <(curl -fsSL https://raw.githubusercontent.com/MSKazemi/yazses/main/install.sh)One command for Ubuntu, Debian, Fedora, Arch and their relatives. It installs YazSes plus the audio, hotkey and typing helpers your session needs, then runs yazses doctor.
brew tap MSKazemi/yazses
brew trust MSKazemi/yazses
brew install --cask yazsesApple Silicon. On an Intel Mac, download the x86_64 .dmg from the releases page instead.
winget install MSKazemi.YazSeswinget can trail the newest release by a few weeks; the installer, Scoop and Chocolatey track every one. The build is not code-signed yet, so SmartScreen asks once: More info → Run anyway.
The Windows guide →pipx install yazsesAny OS with Python 3.11 or newer. On Linux, follow it with yazses setup, which installs the audio and typing helpers.
yazses doctor
yazses mic-level --setIf doctor says you were added to the input group, log out and back in once. mic-level --set tunes the voice gate to your microphone and room.
Open YazSes from Applications (on macOS 15+, the first blocked launch is approved with Open Anyway in Privacy & Security). Then switch it on under Accessibility, Input Monitoring and Microphone, and relaunch it. Input Monitoring is the switch people miss.
Why there are three switches →YazSes starts in the system tray, next to the clock. If Windows asks about the microphone, allow it. The tray icon’s colour shows what it is doing.
What the tray colours mean →yazses doctorChecks the microphone, the hotkey, the typing backend and permissions, and names the fix for anything that is missing.
Something flagged? →yazses startNow hold Right Alt, speak, and release. The words are typed wherever your cursor is, in any app, on X11 or Wayland.
Hold Right Option, speak, and release. The words are typed into whichever app has focus.
Hold Right Ctrl, speak, and release. The words are typed into whichever app has focus.
yazses startHold the hotkey (Right Alt on Linux, Right Option on macOS, Right Ctrl on Windows), speak, and release.
Next: the five-minute guide · voice commands · transcribe a recording · try it without installing
In development
Generated from the changelog on every build, so this section can’t drift from the code. Development is open: every change lands with its tests, and every decision is written down.
Latest release
v2.51.0
v2.51.0 makes gaze status, doctor and status list what Glance-Type can and cannot do on your session, reports a click observer that failed to start, refuses a feature toggle that would half-apply a language profile, keeps Chinese meeting minutes inside the summariser's context, and replaces a stale macOS daemon on reopen.
On main, not yet released
0 changes waiting for the next release
Recent releases
Architecture
Every time you hold the key, your audio flows through a fixed pipeline that never touches the network. Hover or tap a stage — the pulse shows a burst moving through it.
Hold-to-talk starts a burst; releasing ends it. The trigger can also be an EMG muscle sensor over USB serial — a hands-free option.
src/yazses/platform/emg/ The spine is four states. Meeting capture, read-back, the SSH remote path and enrollment are modes you enter deliberately over IPC — and return from.
“What starts a burst” is a contract. A key and an EMG squeeze already implement it; a wake word, a switch or a BCI would be a third — without touching the pipeline.
SttEngine made Parakeet and Moonshine config changes, not a fork. Anything that turns audio into words plugs in here.
Every OS difference is a Protocol. That is what makes Android a port rather than a rewrite.
Where things stand
Four independent summaries, each using the status words of the roadmap. They are not stages: none of them leads to the next, and nothing here is a delivery date. Nothing leaves your machine, and new features start switched off.
Status: Released
Local hold-to-talk dictation, file transcription and opt-in Meeting Mode, with several local speech engines.
Claim limit: released means reachable in a tagged build, not verified on every OS or app.
Status: Released · real-client check pending
The local MCP server (stdio) with ask_human, an opt-in send-to hand-off, and a one-click .mcpb bundle.
Claim limit: pre-fill without sending is the vendors’ documented behaviour, not yet observed here.
Status: Experimental · hardware-unverified
Gaze targeting, head pointer and Voice Pilot, plus EMG, blink and jaw-clench switches. All opt-in and off by default.
Claim limit: real-device evidence is not established; a shipped reader is not a qualified hardware device.
Status: Research · Deferred · not scheduled
Ideas that are open to evaluation or waiting on a decision. None is a promised release.
Claim limit: ADR-030 freezes new capability waves outside its four pillars; lifting that takes a recorded decision with evidence.
Status words, as the roadmap defines them: Released is in a tagged build; Experimental is opt-in with synthetic or CI evidence only; Research and Deferred are not scheduled. What each word does and does not establish
Cloud escalation is fully specified and stays unimplemented, so the offline promise can never erode by increments. Voiceprints, the learning corpus and third-party capture may never leave at all.
ADR-019Third-party plug-ins would sit on the hot path with your microphone, transcript and keyboard. They stay out until a real isolation boundary exists.
ADR-018Personal speech adapters — including for dysarthria, ALS and Parkinson’s — ship only after a measured accuracy win on held-out data. Ideas don’t ship on their own strength.
Future workDocumentation
Guides for the first five minutes, references for the fiftieth hour, and design records for every decision in between.
Install on Linux, macOS or Windows and dictate in minutes.
Wayland, coding by voice, confidential work, accessibility.
Hotkeys, vocabulary, macros, low-RAM models, air-gapped use.
Every command, flag and config key — generated from the code.
The pipeline, the daemon, the platform seams and the figures.
The cited science of post-keyboard input: eye, voice, muscle.
Where it is going, what shipped, and what is declined.
Find a task that fits — many need no code at all.
40-second tour: the core loop, the command line, and the system tray. Terminal output is real; the command-line typing is re-enacted for legibility. Watch it with narration and chapters on YouTube
Fully offline by default & private
Speech is transcribed on-device with CPU faster-whisper (int8). No GPU, no network, no account. No audio, no text — nothing leaves the machine by default.
Types into any app
Hold the hotkey, speak, release — the text lands in whatever window has focus: editor, browser, terminal, chat. Works on X11 and Wayland.
Works over SSH and Remote-SSH
Because text is injected at the OS level, not inside an app, it lands in VS Code / Cursor Remote-SSH panes, integrated terminals, tmux and container shells — where in-app dictation usually can't reach.
Voice commands & macros
A regex grammar maps "undo that", "save file", "go to line 42" to real key sequences.
Transcribe recordings
yazses transcribe meeting.m4a turns any audio/video file into text — offline, with --diarize speaker labels and subtitle export.
Built for accessibility
VAD calibration, mic-level tuning, dysfluency-friendly mode, and an experimental EMG muscle-sensor trigger (off by default; no real-device evidence yet, see the known limits).
Self-improving, on your terms
An opt-in, encrypted, on-device learning corpus lets yazses tune propose accuracy fixes from your own corrections.
The universal installer provisions audio, hotkey and text-injection prerequisites, then runs yazses doctor. See the Linux install guide for APT, Snap and pipx alternatives.
Preview: supported and installable, not yet verified on hardware by a person on a current build. The validation matrix shows what testers have confirmed so far.
On Apple Silicon, Homebrew is the recommended route:
Apple Silicon and Intel also have native .dmg builds. See the macOS install guide for downloads, Gatekeeper and Accessibility/Microphone permissions.
Preview: supported and installable, not yet verified on hardware by a person on a current build. Two testers have reported dictating on Windows 11; the validation matrix shows what has been confirmed.
winget installs the x64 build, which also runs on ARM64 PCs under emulation. It can trail the newest release by a few weeks; the installer on the latest release, Scoop and Chocolatey (choco install yazses) track every release. See the Windows install guide for SmartScreen and microphone permissions.
The snap on Wayland
The strictly confined snap dictates on X11, and on GNOME/KDE Wayland through the xdg-desktop-portal RemoteDesktop API. It cannot configure or use the host ydotoold service, so the portal is the route it takes; the first dictation asks once for permission and the answer is remembered. The unconfined installs above remain the most-tested Wayland path.
On X11 and Wayland alike, connect both required interfaces after installing, then start it:
sudo snap install yazses
sudo snap connect yazses:audio-record
sudo snap connect yazses:raw-input
yazses doctor
yazses start
yazses setup is not required for the snap. Confinement prevents it from installing host packages, changing groups, or configuring host services, so it can only print the manual permission checklist; the snap already bundles its X11 dependencies.
Installed with pipx on Linux? Provision the host in one command. Skip this if you used the universal installer or APT path — both already do it:
yazses setup # installs audio + injection deps, joins the input group, sets up ydotoold
# then log out and back in (the input-group change needs a fresh login)
Then, on Linux or a pipx install:
yazses quickstart # machine-tailored next steps; read-only
yazses doctor # check mic, injection backend and permissions
yazses mic-level --set # tune the silence gate to your room and voice
yazses start # start the dictation daemon
On macOS (Homebrew or .dmg) and Windows (winget or the installer) there is nothing to type: open YazSes, grant the permissions it asks for, and it runs from the menu bar or system tray. Since v2.44.0 the Homebrew cask puts a yazses command on your PATH; with the .dmg, the CLI is /Applications/YazSes.app/Contents/MacOS/YazSes --cli doctor.
Hold the hotkey (Right Alt on Linux, Right Option on macOS, Right Ctrl on Windows), speak, release — the text appears in the focused app within one to two seconds on a laptop CPU. The quick start walks through each OS step by step.
yazses transcribe meeting.m4a turns any audio/video file into text, fully offline. Add --diarize to tag who said what and --format srt for subtitles. See the transcription guide.yazses tune proposes accuracy fixes from your own corrections.YazSes is not an LLM agent — it dictates text and runs editor/terminal commands; it does not browse, reason over your files, or hold a conversation. It uses CPU faster-whisper (a cloud service may still win on raw accuracy for a noisy mic), ships English-tuned *.en models by default, and is desktop-only.
Hold hotkey → record audio → VAD gate → faster-whisper (CPU)
→ clean + disfluency filter → command grammar (Tier 1 regex)
→ dictate? type it · command? send keys
Two at-a-glance signals show you what YazSes is doing: a top-bar "Y" tray icon whose colour is a live state indicator (🔵 idle · 🟢 dictating · 🟡 no text target → clipboard · 🟣 command mode · 🔴 problem) and an optional sonar overlay that pulses near your cursor while it's listening. See Tray icon & overlay for what each colour means.
Want the CLI itself, not a description of it? Watch it play right here — a real asciinema recording of -h → about → quickstart → features → status, with selectable, copy-pasteable text. Or in your own terminal: asciinema play docs/demo/yazses-cli.cast.
Curious why it works this way?
The research section is a public, fully-cited notebook on post-keyboard input: how accurate webcam eye tracking really is, how offline speech recognition overtook the cloud, and why a $50 muscle sensor beats a $1,000 EEG headset. Every design decision in YazSes traces back to a measurement there — and the open questions are open to anyone.
Does it work without internet? Yes — transcription runs locally; nothing is sent anywhere by default.
What GPU do I need? None. It runs on CPU; 4 GB RAM minimum, 8 GB comfortable.
Does it work on Wayland? Yes. Unconfined installs use wtype/ydotool; the strictly confined Snap uses the xdg-desktop-portal RemoteDesktop API on GNOME/KDE Wayland. The universal installer remains the most-tested Linux path. A hardware-verified record on GNOME and KDE Wayland is still to come (#643).
Is it a replacement for Talon? YazSes focuses on offline dictation plus a practical command grammar. Talon has far more advanced scripting. They can coexist.
More answers in the full FAQ, and a side-by-side in Comparison & alternatives.
Apache-2.0 licensed. If YazSes is useful to you, a ⭐ on GitHub helps others find it.
Want to help? See Contributing — several tasks need no code at all, like translating the README, sharing a config you already use, or adding the microphone that worked for you.
Free, open source, and fully offline — on the machine you already own.
Questions or feedback? mohsen.seyedkazemi@gmail.com