Kara v0.3.0

Two beta testers said the app was slow and lost words. Both of them turned out to
be running the smallest model on their processors — one of them on an RTX 3060 —
because the download left the CUDA libraries out and hid the graphics card option
along with them. So the first thing this release fixes is not the speed. It is
that nobody, the developer included, could tell what they were actually running.

Added

• A second installer, for NVIDIA cards. Kara-Setup-GPU carries the CUDA
  libraries, which is why it weighs about 570 MB against 72 MB, and on a card
  it is not a small difference. Nothing else is needed: no CUDA Toolkit, no separate
  download, just a current driver. Either installer can be run over the other and
  they share their settings. AMD and Intel cards are still not supported, because
  ctranslate2 has no backend for them on Windows.
• Numbers are typed as numbers. "mil doscientos pesos" comes out as
  1200 pesos, "son las tres cuarenta y cinco" as son las 3:45, "tres coma cinco"
  as 3,5 and "un veinte por ciento" as un 20%. Spanish and English. A bare "un",
  "una" or "uno" is left alone, because those are articles and pronouns far more
  often than they are the number one. Settings -> TEXT turns it off.
• Spoken punctuation. Say "punto y aparte", "nueva linea", "abre parentesis"
  or "signo de interrogacion" and you get the punctuation rather than the words.
  Only phrases nobody says by accident: plain "coma" and "punto" are ordinary
  Spanish, so they live behind the "all" setting and are off by default.
• Opening question and exclamation marks. Whisper supplies the closing one and
  drops the opener perhaps half the time; a Spanish sentence ending in ? now gets
  its inverted mark back.
• Export diagnostics, in Settings. Writes a zip with a report on the machine
  (processor, graphics card, microphones and their audio backends, resolved
  device and model, library versions), the settings as saved, and the timing log.
  Then it opens the folder and stops: nothing is sent, and the file is yours to
  pass on or delete.

Changed

• The graphics card option now appears whenever a graphics card can be used.
  Until now the packaged build hid it unconditionally and forced Device back to
  auto, because no packaged build could use a card. That was true, and it was
  also why two beta testers spent an evening reporting that the app was slow and
  lost words while both of them ran the smallest model on their processors -- one
  of them on an RTX 3060, with everyone involved, the developer included,
  believing the card was in use. The three places that inferred this from "is
  this a packaged build" now ask whether the CUDA libraries load.

• Auto no longer refuses large-v3-turbo unless 6 GB of video memory is free.
  Its weights are about 1.6 GB, so the old threshold was several times what it
  needs, and the effect was that any card with a browser open fell all the way
  to small. The scale is now 2.5 GB for large-v3-turbo and 1.2 GB for
  small. There is deliberately no medium step between them: it wants as much
  memory as large-v3-turbo while being both slower and less accurate, so
  anywhere medium would fit, the better model fits too.

• The timing log now runs always, rather than only when the KARA_DEBUG_LOG
  environment variable was set. Nobody was ever going to run setx before using
  a dictation tool, which meant the machines that were slow were exactly the ones
  that had recorded nothing. It is capped at 2 MB, it still records timings and
  device names and never what was said, and Kara still has no way to send it.
  KARA_DEBUG_LOG is ignored from now on, since there is nothing left for it
  to switch on; KARA_MACHINE still names the machine inside the file.

• ctranslate2 gets one thread per physical core, up to eight, instead of the
  four it defaults to regardless of the machine.

• Beam width is 1 on the processor and stays at 5 on a graphics card. Beam 5
  weighs five candidate transcriptions for what is usually a comma's worth of
  difference, and on a processor that is most of the wait.

• Silence is trimmed before transcription. Push-to-talk recordings start and end
  with some, because the key goes down before you speak and up after you stop.

  Together those three take a 66-second recording on a six-core processor from
  14.8s to 9.1s, with byte-identical output on the test sample.

• Whisper is now primed with a formatted sentence before long recordings, on
  the theory that it copies the punctuation style of whatever it is primed with.
  Said plainly: this could not be shown to help. On 66 seconds of clean, well
  articulated speech the output was byte-identical with and without it. The case
  it is aimed at is fast unpunctuated speech, which is the one case a synthesised
  test voice cannot produce, so it ships unproven. It is skipped on recordings
  under three seconds, where it was measured steering the result somewhere worse
  and where there is no punctuation to recover anyway.

Fixed

• Dictating no longer wipes the clipboard when it held an image or a copied
  file. The old code read the clipboard as text, got an empty string back, and
  wrote that empty string over the picture once the paste was done.

---

Download: Kara-Setup-0.3.0.exe — Windows 10 and 11.

Windows will show a blue "Windows protected your PC" box, because the installer is not signed. Click More info, then Run anyway. The install guide walks through it and shows how to check the file's SHA256 against SHA256SUMS.txt below.