gsuberland, I was looking into Codec 2 for low bitrate telephony, and it turns out some researchers made their own decoder for it called Parametric WaveNet, which is a deep learning model that generates speech directly from the Codec 2 data stream.
listen to the Codec 2 samples, then the Parametric WaveNet samples. they're both the exact same encoder at 2400bps. the difference in quality and inteligibility is outstanding.
https://storage.googleapis.com/downloads.webmproject.org/icassp2018/index.html