Skip to content

v2.43.2 — no more words nobody said

Released 2026-10-18. Upgrade:

sudo snap refresh yazses     # snap
pipx upgrade yazses          # or: uv tool upgrade yazses
scoop update yazses          # Windows

Restart the daemon afterwards (yazses restart). A newly started YazSes replaces an older running one since v2.43.1.

A loud burst with no speech in it is not typed

Press the hotkey in a room with a fan, or close your eyes for the face switch without saying anything, and YazSes could type "The app is called YazSes." (#598). That sentence is the name hint the speech model is primed with. Given a burst with no speech in it, the model does not answer with nothing: it answers with the words it was primed with. This reproduced on synthetic noise (no one's audio) with the small, base.en and tiny.en models, and a personal vocabulary came back the same way. The volume gate cannot tell a talker from a fan, and the model's own "no speech" score is too close between the two to use.

Before decoding, the daemon now asks the speech detector that yazses verify and file transcription already use, and drops a burst only when it positively finds no speech. The log says No speech in a 1.2s burst … -- discarding. with the length and level only, never audio or text. Saying the sentence still types it: the rule looks at the audio, not the words.

What was measured, on a public-domain speech sample cut into 22 one-word bursts and on 27 seeded noise bursts: no voiced word was lost at normal or one-tenth volume, and no noise got through. A crude stand-in for whispering lost 3 of 22 words, so whispering is the weak spot; real whispered, accented and dysarthric speech was not measured. The check is skipped in Whisper / Low-Effort mode, Dysfluency-Friendly mode and the sotto-voce command channel, does nothing if the detector cannot run, and can be switched off with [accessibility] speech_presence_check = false. See silent audio discarding if quiet speech is being dropped.

Eyes-closed dictation: three fixes

  • A stray hold-release no longer re-types the previous dictation. In the first live run one closure was typed and two more releases with no recording typed the same clip again.
  • Binding "eyes closed" in Settings → Hands-free control no longer dictates on every blink: it now also raises the hold before any face switch counts, from 0.2 s to 0.8 s.
  • With hands-free safety on, a face switch whose every binding is a hold re-arms itself once your face has been back for a second, instead of staying off until Resume after a glance at the keyboard.

Smaller fixes

  • The tray supervisor's relaunch line now says what tray.lock looked like, so a "tray is not running" line next to an old error can be read correctly (#578).
  • The gaze grounding resolver no longer grounds a candidate with no extent (a height so small it rounds to zero).
  • The release check no longer marks a release incomplete for having only one Windows installer; v2.43.1 had been demoted to a pre-release by it and was restored by hand.
  • Settings windows: face-switch mapping boxes are no longer cut off, & in category titles is shown, and "resume from the tray" names a menu item that exists.

The complete list is in the changelog.