YazSes

YazSes — Type with your voice. Keep your voice on your device.

Free, open-source, offline voice typing for Linux, macOS and Windows.

Install YazSes Try Hosted Demo View Source on GitHub

The hosted demo runs on Hugging Face servers, so audio you send to it leaves your computer. See Privacy.

YazSes settings window on the Features tab. Dictation core and voice commands are switched on, and optional features such as voice punctuation are switched off.
The Features tab: what is on by default, and what you choose to turn on.

How it works

  1. Hold a key

    Hold your hotkey: Right Alt on Linux, Right Option on macOS, Right Ctrl on Windows. You can change it with yazses hotkey set.

  2. Speak

    By default, audio is recorded only while you hold the key. Speech is recognised on your CPU with faster-whisper.

  3. Release

    The text is typed into the focused app. Commands such as "undo that" or "go to line 42" run as editor or terminal actions instead.

Features

Each feature below is either on from the start or waiting for you to turn it on. Experimental and planned features are not listed.

On by default

  • Hold-to-talk dictation. Always on. Hold, speak, release.
  • Voice commands. "Undo that", "save file", "go to line 42", "run the tests".
  • Mid-thought undo. Say "scratch that" to drop the last phrase.
  • System-tray icon. Pick a microphone, re-calibrate, or start and stop the daemon.

Optional, off until you enable them

  • Transcribe recordings. yazses transcribe turns audio or video into text. Enable with yazses features enable recimport.
  • Meeting capture. yazses meeting start and stop. Speaker labels need the diarization extra.
  • Dysfluency-friendly mode. Collapses stutters and repeats, for example "b-b-because" becomes "because".
  • Voice punctuation. Say "comma" or "new line" to insert marks.
  • Other speech engines. NVIDIA Parakeet and Moonshine, both English, as alternatives to Whisper.

Every switch is listed with yazses features. Each optional feature is turned on with yazses features enable <name>, then applied with yazses restart.

Privacy

There are two ways to use speech recognition from YazSes. They have different privacy properties, so keep them apart.

Option Where speech is processed Does audio leave your computer?
YazSes installed on your computer Your own CPU, with faster-whisper. The speech model is downloaded once, on first use. No. By default, no audio, transcripts, editor context or usage data leave the machine. Dictation works without a network connection after the model is downloaded.
Hosted demo on Hugging Face Hugging Face infrastructure. Yes. Audio you submit to the demo is processed on Hugging Face servers.

The hosted demo lets you judge accuracy before you install. For private, offline voice typing, install YazSes on your own computer. The full statement, with every network connection listed and its trigger, is on the privacy statement page.

Install

Choose your operating system. Each guide gives the commands and the first-run steps.

Linux has the most reports from real machines. The macOS and Windows builds are newer, and the validation matrix lists exactly what has been confirmed on real hardware. Not ready to install? You can try it without installing, or read what installing actually costs first.

Research

The paper

YazSes: An Offline, Privacy-First, Cross-Platform Hold-to-Talk Voice-Dictation System, by Mohsen Seyedkazemi Ardebili.

arXiv:2607.28878. https://arxiv.org/abs/2607.28878

Benchmarks and reproduction

With the default base.en model, the paper reports a word error rate of 4.07 % on a 200-utterance LibriSpeech test-clean subset, with a median decode of 1.56 s on a laptop CPU. Dictation on your own microphone will differ.

Benchmarks and method · Results manifest on GitHub

Community

YazSes is open source. You do not need a speech or AI background to help.