terry matula

https://github.com/matula/ai-ncoder

I hacked this together months ago because I found myself converting a bunch of old audio files, and I was continually googling #ffmpeg flags. I thought it would be cool to use natural language for transcoding, but wanted it to keep it local, and workable offline.

So I grabbed the smallest, usable model and hacked together some Python and prompts... And it just kind of worked.

I thought I'd clean it up and share.

GitHub - matula/ai-ncoder: AI-ncoder

AI-ncoder. Contribute to matula/ai-ncoder development by creating an account on GitHub.

GitHub
Mar 10 at 6:47amWeb
Show threadterry matulaMar 10

99% of my usage is "turn these files into 320k mp3s" or "convert this file to an mp4"... And that's solid. But I also noticed it'll attempt other ffmpeg flags as well, like being able to edit the length of a file... Or convert to gif. I added some volume level shots as well, but haven't really tested it.

I'm not really sure what ffmpeg the Qwen model has by default. But might be cool to custom train one and get the model size as lean as possible.