Skip to content

Releases: kittcat-lab/dsh-kitt-voice

Release list

1.3.0 — the install no longer fails

1.3.0 — the install no longer fails Latest
Latest

Choose a tag to compare

View all tags
@kittcat-lab kittcat-lab released this 09 Sep 15:38
v1.3.0
f8289dc

The install no longer fails, and the plugin no longer arrives switched off.

Measured on a clean machine against the current harness: dsh plugin --profile web add dsh-kitt-voice ended in an error, and because it ended in an error the harness never reached the step that activates the plugin. It was installed and inert. Anyone following the first line of the README saw a failure and then nothing at all. It had been that way since the first published version.

Fixed

  • One dependency was the cause. msedge-tts ships a preinstall of npx only-allow pnpm. A harness profile installs with pnpm, and pnpm refuses to run anybody's install scripts without permission, so the command failed with ERR_PNPM_IGNORED_BUILDS. The Microsoft read-aloud protocol is now implemented in the plugin itself: no install hooks, and the package went from six dependencies to two.
  • The false compatibility warning is gone. Cordis, the web app and React are provided by the harness at runtime and are never installed as packages, so declaring them as required peers made every install report them as missing.
  • Text is escaped before it enters the SSML document. An & or a < in a reply broke the request outright — and an agent's reply is exactly where those appear.
  • Reading speed now reaches the neural voices. It was sent as a bare number instead of the documented percentage. At double speed the audio is now half the size.

Also in this line of releases

  • 1.2.0 — A Piper voice no longer says "jotason" where the text says JSON. Piper reads phonemes generated for its own language, so English terms are rewritten the way a Spanish speaker would spell them, before synthesis and only for Piper. Forty-five entries, editable in Settings.
  • 1.1.0 — When the microphone is the problem, the plugin says so: muted, far too quiet or clipping, with what to do about it, instead of "could not transcribe". It stays quiet when it could not measure the level.

The trade, stated plainly

The read-aloud protocol is ours now, so its drift is ours too. If Microsoft rotates the client version it expects, neural voices stop answering — with a reason on screen, and with Piper and the system voice carrying on. The constant to update is named in lib/neural.js.

Twenty-seven traps paid for and written down: DOCUMENTACION/TRAPS.md

Assets 2
Loading