Unitra AI Desktop 0.11.0

Engine 2 arrives as the Standard recognition model, with 15 more languages you can speak, an interface available in 35, and a round of fixes.

Engine 2 is the Standard recognition model

Open Settings > General > Recognition Model. There are two options:

  • Standard is Unitra Engine 2, the new engine. It is the default on any machine that can run it.
  • Legacy is the engine every previous version used.

If you have read about Engine 2 and are looking for a switch by that name, Standard is it. There is nothing else to turn on.

What it changes day to day:

  • Sentences that mix two languages stay whole. The old engine could only work in one language at a time, so switching language mid sentence dropped the other half or turned it into the wrong word. Engine 2 keeps both.
  • Game vocabulary comes through as itself. Callouts like HP, Rush B and site are recognised as terms instead of unrelated words.
  • Speech holds up better in a noisy match, including over gunfire.

Speech recognition runs on your own computer, and Engine 2 works across NVIDIA, AMD and Intel graphics.

What it needs. Engine 2 uses about 2 GB of graphics memory and needs 3 GB free to start. Machines with less stay on Legacy automatically, with the reason shown in Settings, instead of switching and then failing.

For how Engine 2 performs and how we measured it, see Meet Unitra Engine 2.

15 more languages

The language picker goes from 20 to 35. The new entries are Cantonese, Bahasa Indonesia, Thai, Turkish, Hindi, Bahasa Melayu, Swedish, Danish, Finnish, Czech, Filipino, Greek, Hungarian, Macedonian and Romanian. All 15 are recognised by Engine 2, so they need Standard.

Each of them also comes with a fully translated interface, so the app itself is now available in 35 languages, not just able to hear them.

Gaming vocabulary

Recognition of game terms is better, and it now switches on when you turn gaming mode on rather than waiting for Unitra AI to recognise your specific game. Detection of fullscreen games without a profile improved as well. Gaming vocabulary also updates from our servers now, so new terms no longer wait for an app update.

Fixes

  • Subtitles that did not appear at all, even though the speech was picked up. This included alt-tabbing out of a game being mistaken for closing it.
  • Sentences that shortened after they appeared, on Japanese, Korean, Russian and Vietnamese. A finished subtitle could collapse into a fragment a moment later.
  • Subtitles that ran on during audio with no pauses in it, such as a stream with music playing underneath.
  • Selecting a game manually now applies to how audio is captured, not just to the game name.
  • Keyboard shortcuts. Single key shortcuts such as F5 now work, and a shortcut you set is kept after a restart.
  • Interface language. Auto Detect no longer falls back to English for languages the app already includes, and the overlay follows a language change without needing a restart.

Plus a number of smaller fixes to recognition, audio handling and stability.

Korean

Korean is now handled by Engine 2 only. The Korean model has been removed from Legacy.

Open source licenses

Settings > Account > About > Open Source Licenses now lists the third party models and components inside Unitra AI, with the attributions their licenses require.

If something does not work as expected after updating, let us know in Discord.

← All posts