Skip to content

Ask for the language you speak, not the one Windows is in - #38

Merged
savvasioakeim merged 1 commit into
mainfrom
feat/dictation-language
Oct 31, 2026
Merged

savvasioakeim merged 1 commit into
mainfrom
feat/dictation-language

Conversation

@savvasioakeim

Copy link
Copy Markdown
Owner

Dictation asked the speech service for navigator.language. On an English Windows that means English recognition, whatever the person actually speaks — and the service answers the question it was asked. A Greek sentence requested as English comes back as English spelling of Greek sounds, which looks like a broken recogniser rather than a wrong request.

Measured today on Runtime 155: with the language named, Greek dictation is accurate and immediate.

What changed

  • A language setting, sent as a region-qualified tag (el-GR, not el, which the service rejects). Same as this computer keeps the previous behaviour.
  • The engine switch is gone. Its Whisper option had no backend behind it, so it offered a choice that did not exist.
  • The network message no longer blames WebView2. It asserted a known fault from Runtime 153 on. That is not what happens here: on Runtime 155, dictation works in the dev build and the installed 0.4.0. The claim was wrong, and it sent people to check a connection that was never at fault.

Verified

  • tsc --noEmit clean, npm run build passes
  • Greek and English dictation both tried by hand through the UI, in the running app

Not done

Local recognition. #37 built a full Whisper engine; measured on this machine it reached 0.67× realtime at tiny and 1.35× at base, with Greek only becoming usable at small, which extrapolates past 4× realtime. The quality worth having sits beyond the point where the speed collapses, so that PR is closed rather than merged.

🤖 Generated with Claude Code

Dictation requested `navigator.language`, so an English Windows asked the
speech service for English recognition no matter what was actually spoken.
The service answers the question it was asked: a Greek sentence requested
as English comes back as English spelling of Greek sounds, which reads as a
broken recogniser rather than a wrong request.

The language is now a setting, applied as a region-qualified tag because
the service rejects a bare "el". "Same as this computer" keeps the old
behaviour for anyone whose machine already matches their speech.

Two things go with it. The engine switch is gone: its "Whisper" option
never had a backend and the choice it offered was not real. And the
`network` error no longer asserts a WebView2 fault from Runtime 153 —
tested on Runtime 155 today, dictation works in both the dev and the
installed build, so that claim was wrong and sent people to check a wifi
connection that was never the problem.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@savvasioakeim
savvasioakeim deleted the branch feat/qa-capture-and-dictation October 31, 2026 18:29
@github-actions github-actions Bot locked and limited conversation to collaborators Oct 31, 2026
@savvasioakeim savvasioakeim reopened this Oct 31, 2026
@savvasioakeim
savvasioakeim changed the base branch from feat/qa-capture-and-dictation to main October 31, 2026 18:29
@savvasioakeim
savvasioakeim merged commit 4786443 into main Oct 31, 2026
4 of 6 checks passed
@savvasioakeim
savvasioakeim deleted the feat/dictation-language branch October 31, 2026 18:49
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant