Home · Projects · YazSes

Active · open source · Apache-2.0

YazSes — your voice, typed on your own machine.

Hold a key, speak, release — and the words appear in whatever window has focus. The speech never reaches a server: it is transcribed on your own CPU, so privacy is a property of the architecture rather than a promise in a policy. Nothing leaves the machine by default; there is no account and no API key.

Documentation & install Source on GitHub
Active Project · Privacy & accessibility

What it does

YazSes is a voice human–computer interaction layer, not only a dictation box. It does three things, all on your own CPU:

  • Dictate — hold a key, speak, release; the text is typed into the focused editor, browser, terminal or chat window. Spoken editor and terminal commands (“undo that”, “go to line 42”) are recognised as commands rather than typed as text.
  • Transcribe a file — yazses transcribe interview.m4a returns a transcript of any audio or video file, optionally tagged with who said what, as txt, md, srt, vtt or json.
  • Capture a meeting — yazses meeting start produces a speaker-labelled transcript and, optionally, minutes written by a local language model. Speaker names come from voiceprints you enroll yourself, never from a cloud account.
pipx install yazses && yazses quickstart
Why this is a Responsible Systems Lab project

Responsibility here is an architectural property, not a policy page.

Dictation is one of the most intimate things software can touch. To use most of it, you send a recording of your own voice — a medical conversation, a legal matter, a confidential meeting, a first draft you are not ready to show anyone — to a company's servers, and then trust a privacy policy about what happens next.

YazSes removes the need for that trust rather than asking for it. There is no network call in the dictation path, so there is no server to be breached, no retention policy to change, no account to be sold with the company. That is the distinction this lab cares about: a guarantee you can verify by reading the source and watching your own network interface, instead of a guarantee you have to take on faith.

Accessibility is part of the same commitment

Two features exist for people that mainstream dictation tends to serve badly:

  • Dysfluency-Friendly Mode — an opt-in collapse of stutters and repeats (b-b-because → because) for stuttered or dysarthric speech, which most speech-to-text transcribes literally or discards.
  • EMG trigger support — YazSes can be triggered by a USB muscle sensor rather than a keypress, for hands-free use by people with motor disabilities.

Neither feature enlarges the addressable market much. Both are the reason the project is worth doing.

Stated limits

What YazSes deliberately is not.

Being specific about what a system cannot do is part of what makes claims about what it can do worth anything. YazSes is not recommended if you need:

  • A conversational AI agent. YazSes dictates text and runs editor and terminal voice commands. It does not browse, reason over your files, or hold a conversation. An agentic version was prototyped in an archived branch and is not shipped.
  • Non-English languages out of the box. The default model is English-only. Other languages need a multilingual model and explicit configuration, and accuracy is honestly described as usable but still rough.
  • A mobile or web app. It is a desktop daemon. An Android implementation is in design, not released.
  • The lowest possible word-error rate on a noisy microphone. A cloud speech service may still beat CPU faster-whisper there. The trade-off is that nothing leaves your machine.
Of the 141 capabilities YazSes documents, 74 are wired and 67 are marked “planned” in public rather than described as if they already worked.

The same discipline applies to the privacy claim itself. Dictation makes no network call and writes no audio to disk. Meeting capture is different: it writes a recording to disk while it runs, and deletes it after transcription unless you ask to keep it. Both stay on your machine — but they are not the same claim, and this page will not flatten them into one.

Use it, or help build it.

YazSes is free and open source under Apache-2.0, maintained in the open with 12 contributors. Issues, fixes, translations and platform testing are all welcome.

Open the repository