I am aware of lyra. Its pretty good. Very computationally expensive though - no audio application I have ever worked on had even close to the power/thermal budgets that would allow use of the deep-learning codecs. Maybe someday we will get very low-energy hardware accelerators for them, but until then, these are a non-starter for things I work on.

The thing is (and maybe this is a nitpick), once you are down to several hundred bps for speech, its getting to be more like speech-to-text (the encoder) and text-to-speech (decoder) than an audio codec.

I am actually not aware of any non-speech audio codecs which can go that low. Any links?

When we have anough compute or the models get much smaller. At this point it seems wasteful to utilize gaming level GPU to decode audio stream.

Ah yes, should have guessed. WMA was pretty big in the early 2000s. Around 2004 I even had access to the official wma source from Microsoft.

Spent about 4 months porting and then optimizing the wma decoder on a proprietary very-low-power dsp. It was hard work, but I got it down to about 12mhz for real-time decode on the dsp. The product this was intended for was going to be powered by a watch battery - hence the extreme optimization requirement. The source code, as received, used to consume ~400mhz on an athlon iirc. But with a lot of assistance from the dsp hardware and 4 months of elbow grease it could decode in less than 12mhz - for some profiles. Insane codebase!