OpenWord v1.22.0

Added

• The licences, in the app. Settings credits the voice — KittenTTS,
  Apache-2.0; sherpa-onnx and ONNX Runtime; and eSpeak NG, GPL-3.0, with a
  link to the source of the version built in — and a new Licences entry
  shows the full text of every licence, the voice's among them. eSpeak NG
  is compiled into the library the voice runs on, so the app as built is
  distributed under the GPL-3.0 as well as the MIT licence of its code.

Changed

• One voice, built in. Reading aloud is now always in KittenTTS, the
  open neural voice, which ships inside the app: nothing to download, and
  Listen works straight after installing, with no connection. Its eight
  speakers are chosen in the voice sheet. Kokoro and the device's own
  voices are gone, and with them the flutter_tts plugin; a voice chosen from
  either before is read as Kitten's first. The files 1.21.0 downloaded are
  deleted. The download is about 55 MB larger for it.
• Kitten is at full precision. The 8-bit edition 1.21.0 offered now and
  then made a garbled sound — KittenML itself warns of it — and was, on the
  processors tried, three times slower to make speech than the full one.
• Reading aloud on Linux, which it never did before, since the voice no
  longer comes from a plugin Linux lacks. Not on the web any longer: a
  browser cannot run the voice.

Fixed

• OpenWord's voice took a long time to start, and paused part way.
  Its model was loaded only when Listen was tapped, and a model's first
  run is far slower than any after it; it is now loaded, and run once, in
  the background as the app opens for anyone who has listened before.
  Nothing was heard until the first piece had been made whole, and that
  could be a long sentence; the first piece is now forty characters at
  most, cut where the words allow, and each after it may be twice as long,
  so the voice starts almost at once and keeps ahead. Reading was handed
  over a passage at a time, each starting from nothing, which was a
  silence every minute or so; the voice now has the rest of the chapter at
  once, and three pieces are made ahead instead of one. Where a device
  still cannot keep up, the reader says so once.
• A crackle on the loudest syllables. The voice's output can go past
  full scale, and the loudest samples were cut off flat as they were written
  out. A piece that would be is now made a little quieter as a whole.

---

Every translation is bundled with the app — the World English Bible, the
Berean Standard Bible and the World English Bible, British Edition, all public
domain — along with the book introductions and the maps. Nothing is
downloaded to read it.

From 1.5.0 the app does one thing over the network: it asks GitHub, once a
day, whether a newer release is out, and on Android can fetch and install it
for you. That is the only reason it now asks for the internet and
install-packages permissions, and it can be switched off in Settings, after
which nothing reaches the network at all.

Install the Android APK with adb install OpenWord-*-android.apk, or allow
installation from unknown sources on the device. It is not published on Play.

iOS and macOS builds are not attached because they need an Apple signing
identity; build them yourself with flutter build ipa / flutter build macos.