I wonder if there could be an option to share the generated ai subtitiles so there is no duplicate effort. For example Movie_torrent.vlc subs can be shared on some other users machine with the same Movie_torrent.vlc file.

Obv you can find majority of subs online for popular files but niche cases usually have low subs available and sharing AI could help the ecosystem.

Given modern mirrorless cameras can detect and track humans or mammals using DSPs and NNs without any discernible battery penalty at 30/60/120/240FPS (depending on camera class), I think they can do it with very small and efficient models with some frequency clamping and downsampling.

Initial versions will be more power hungry, but it'll be negligible at the end, given that modern processors have accelerators for image and voice based AI applications.

Whisper.cpp ran fine on my GPU-less 2013 era 8GB i5 Dell laptop.