New AI compression algorithm called SOTA: 10x better compression than MP3

[pierre]

This blog describe how using AI Meta researcher can compress 10x better than MP3. They made the code open source so you can play with it.
Can you hear a difference?

 

[MCH]

It is indeed impressive compared with the two others and in absolute terms if you think it is 6kb/s, but the differences with the original are obvious. The high pitch bell in the original disappears completely in the compressed version and it seems there is very audible distortion in the second part of the clip.

 

[RandomEar]

Google did something similar last year with Lyra (ultra low bitrate for speech) and SoundStream (multi purpose). Looks like the future is now

 

[Curvature]

I think this is not meant to replace two channel recordings, but to allow streaming of complex multichannel recordings, ambisonics and the like, with 10s of channels, for immersive environments.

 

[charleski]

In installed the compressor (like most OS stuff, this was a matter of hunting down the correct package manager to let you install the right dependencies and then hope you had the environment configured right...).

Since Meta's examples page only goes up to 12kbps I thought I'd see how well this codec does at its maximum settings. Below are links to a fragment of Beethoven's last string quartet encoded with -b 24.0 --hq (the input was first resampled to 48kHz in Izotope RX). Encoding the full 7m 25s track took over 324 seconds using my wimpy GTX1060.

Meta Compressor - Google Drive

drive.google.com

I think the difference is clear enough (even to a cloth-eared barbarian like me) that it's not worth trying to ABX them. The result is reasonably good, but it reminds me of trying to listen to music on a medium-wave AM radio ... so might appeal to those into retro gear .