substitcher202309 .zip 14MiB

requirements

pip3 install pysubs2 titlecase

sudo apt install jq kid3-cli rename ffmpeg

flatpak install flathub org.freac.freac

Purpose is to identify hallucinations, repeating subs, stuck timecodes, repeating timecodes.
Biggest difference I've noticed is medium is better to not hallucinate than medium.en and large (which is largev2). I've also tried quantized 5 model but accuracy is as bad as small so might as well just use small in that case.

Substicher comes with a sample librivox audiobook to quickly play around with the options. You put all your srt or vtt subs into the root directory along with the opus audio segments and stitch them together. With the included audiobook extract by chapters and rename 001.opus, 002.opus which is option h) and that will correspond to the whisper.cpp transcribed vtt or srt.

Play audiobooks with subs with a black cover image. Linux SMPlayer, Windows PotPlayer, Mac IINA, Android mpv-android, iOS $ nPlayer or Liquid Player.