YazSes

Offline voice typing for Linux, macOS and Windows. Hold a key, speak, release. The words are typed into whatever app you are using.

YazSes runs speech recognition on your own computer, so your voice does not leave it. It is free and open source under the Apache-2.0 licence. It needs no account, no subscription and no API key.

What it does

• Type with your voice (always on). Hold the hotkey, speak, release. The text goes into the window that has focus.
• Voice commands (on by default). Say "undo that", "go to line 42" or "run the tests" and YazSes carries out the action instead of typing the words.
• Recording transcription (optional, off by default). yazses transcribe turns an audio or video file into text. Enable it with yazses features enable recimport.
• Meeting capture (optional, off by default). yazses meeting start and yazses meeting stop record a meeting and write a transcript. Speaker labels need the optional diarization extra.
• Other optional features, all off until you turn them on: voice punctuation, a voice-activity overlay, an encrypted local learning loop, and the NVIDIA Parakeet and Moonshine speech engines as alternatives to Whisper.

Dictation works out of the box. Each optional feature is switched on with yazses features enable <name> and applied with yazses restart.

Limits. The default speech models are English (*.en). YazSes is a desktop program. It is not a chatbot or an AI agent, and it has no mobile or web app.

Privacy, in plain words

• Speech recognition uses faster-whisper on your CPU. By default, no audio, transcripts, editor context or usage data leave your computer.
• The speech model is downloaded once from Hugging Face the first time you use it. After that, dictation needs no network connection.
• Two opt-in features can send text you dictated: yazses remote <host>, and dictation cleanup through Ollama. Both are off by default.

Read the full privacy statement.

Hosted demo: a different privacy setting

Open the hosted demo to try the accuracy of the same engine and model in a browser.

This is a hosted demonstration. Audio submitted here is processed on Hugging Face infrastructure. For private, offline voice typing, install YazSes on your own computer.

Platforms and how far they have been verified

The project's validation matrix lists what has been confirmed on real hardware, with the evidence for each entry. The maintainer does not own a Mac or a Windows PC, so results for those systems come from volunteers. A short report from a real machine is one of the most useful contributions you can make.

Install

• Linux guide
• macOS guide
• Windows guide
• Try it without installing, in Docker or the browser
• What installing actually costs, before you start
• Python package: yazses on PyPI

Research

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. With the default base.en model, the paper and the benchmarks page report a word error rate of 4.07 % on a 200-utterance subset of LibriSpeech test-clean. Median decode time was 1.56 s on a 13th-generation Intel Core i7 laptop CPU, with int8 inference and no GPU. Read speech in a clean recording is easier than dictation on your own microphone, so expect your own results to differ.

Reproduce. The benchmark commands and the method are on the benchmarks page. Every archived measurement, including runs that were later superseded, is listed in the results manifest.

Models used (third-party, not trained by YazSes)

YazSes does not train speech models. It uses these open models:

• faster-whisper speech models by SYSTRAN: tiny.en, base.en (default) and small.en (tiny.en, base.en, small.en).
• Optional engines: NVIDIA Parakeet TDT (parakeet-tdt-0.6b-v2, English) and Moonshine (moonshine/base, English). Each is off until you enable it.

Each model's own licence and terms apply to that model.

Contribute

YazSes is open source and welcomes contributions that need no speech or AI background:

• Test it on your computer and report what happened.
• Open issues
• Good first issues
• Source code

If YazSes is useful to you, a star on GitHub helps other people find it.

Licence and attribution

YazSes is released under the Apache License 2.0. See the NOTICE file for attribution. Documentation: mskazemi.com/yazses.