As #612 is closed, I open a new one, as I had the same problem. I'm using the latest main branch from 2024-10-02 with the large-v3 model. I played around with the parameters and found the following:
I had some success with playing around on the context. Sometimes -mc 50 worked OKish, -mc 100 gives better quality, but -mc 1000 fails often on my audio.
What mostly worked on all my files (4h of conference files with speakers and round-tables, all in English) so far is
-mc 50 -bs 6 -bo 6 -tp 0.2
But the quality is definitely lower than with -mc 0. Except that it doesn't have that many repetitions... Some were still there :(
As #612 is closed, I open a new one, as I had the same problem. I'm using the latest
mainbranch from 2024-10-02 with thelarge-v3model. I played around with the parameters and found the following:I had some success with playing around on the context. Sometimes
-mc 50worked OKish,-mc 100gives better quality, but-mc 1000fails often on my audio.What mostly worked on all my files (4h of conference files with speakers and round-tables, all in English) so far is
-mc 50 -bs 6 -bo 6 -tp 0.2But the quality is definitely lower than with -mc 0. Except that it doesn't have that many repetitions... Some were still there :(