kanjieater
commented
1 year ago I went ahead and implemented all 3. I can't say it's in a state you would want to merge it back in, as I broke whatever the flac demuxing part was doing in favor of not writing an extra file, and just sticking with the source input. I also am ignoring all the things like bitrate, normalize, and to_mono. You could reenable them, but I figured it would be faster for splitting files to do it this way.
So this is really only focused on the srs command. But the results are astounding.
On my i9-13900ks, 32 core (obviously a bit of an edge case)
1st run, defaults 6hr m4b audiobook 165mins, 800mb of audio
2nd run, Bitrate changes & minimizing re-encode step resulted in 105mins, 86mb
3rd run, parallelize across core count 6mins
Final result 6mins, 86mb
You can check it out here. https://github.com/kanjieater/subs2cia
Feel free to reincorporate anything I may have commented out - I think this is good enough for my use case now.
I'm using audiobooks with synchronized texts, so this sometimes uses large files. The current 6hr .m4b file is only around 150mbs. It is ~57k bitrate for audio quality.
The srs command has been running two hours to process this. The things I noticed immediately are:
The bitrate should probably be preserved from the source instead of defaulting to 320k. I think ffmpeg-python does this by default, but it's explicitly stated. This will cause your files to be bigger for no improvement in audio quality.
Use multithreading for the conversion. I have a 32 core cpu, that is only using about 4~10% for this task. Ideally we'd max that out at 100%. I'm not sure if ffmpeg-python supports --threads command out of the box that ffmpeg supports or if that would actually improve performance. The alternative would be to chunk up the files that aren't re-encoded into enough parts to keep each cpu busy.
Option to not reencode. It looks like there isn't a way right now to just keep the given format and just split it by subtitle timestamp, and output the csv. This would probably be the fastest of all options. Maybe this already exists by encoding the options properly, but I couldn't figure it out.
I'm happy to break these out into other options. If there's one of these I could contribute on just let me know.