Skip to content

0.35.0

Choose a tag to compare

@mikwiseman mikwiseman released this 27 Oct 14:33
· 23 commits to main since this release
9bd92b2

OpenRamble 0.35.0

A third of the memory

  • OpenRamble now runs the speech engine it was measured with, transcribe.cpp 0.4.1. Until now the app was running an older copy by mistake.
  • It needs about a third of the memory: about 270 MB once the model is loaded instead of about 740 MB, and at most about 370 MB instead of 1.1 GB. The model is read straight from its file, so when your Mac runs short of memory it can take that memory back and read it again later, instead of writing it to swap.
  • Text lands 25 to 35% sooner after you release the key, both on a quiet Mac and with every core busy. The model loads in 0.07 s instead of 0.16 s.

More accurate

  • Fewer errors in dictation: from 5.5% to 4.8% in Russian and from 5.1% to 4.9% in English on a standard test set.
  • Long recordings, such as imported files and dictations over 30 seconds, are now recognized in 25-second pieces, because the new engine can silently drop speech from longer ones. Long Russian recordings come out more accurate. Long English recordings make about half a point more errors than in 0.34, and recordings that mix languages up to two points more; the next release works on that.

Your voice and text still never leave your Mac.