Your info about AMD’s addition of AVX2 not until 2017 raised the priority of this for me. I found an acceleration library to add to whisper.cpp that does runtime CPU detection and turned off the usage of AVX2 in the whisper.cpp build itself. On my system this is actually faster than what we had in the beta! Please check tomorrow’s build.
3 Likes
Great work, @shotcut. Awesome and very clever.
Working !!!
The acronym RIM (Research In Motion) is spelled incorrectly, but to be fair, when Matthias Wendel says RIM, it kind of sounds like REM. So no points deducted for that mistake 
Thanks a lot @shotcut for finding this fix so quickly. 
1 Like
Question, do you guys have progress for the speech to text job? For me it’s just --;–;-- but if I view the log I can see the last timestamp and compare to the project’s length to get an estimate.
The latest nightly is also slightly faster for me (13m37s vs 15m20s for a 40min video).
I notice this, too. The progress indicator from whisper.cpp is not fine-grained. And it changes depending on the number of threads running (which depends on the number of cores in your system). On shorter clips, I never even see a progress display. On longer clips, it seems to “jump” in chunks rather than displaying a consistent progress.