Reminds me of Xiph's Speex/CELP model of speech as a mix of noise and frequency to achieve high compression, requiring as little as 2.15 kilobits (275 bytes) per second. It sounds perceptibly similar to the original recording, even though the difference between the input and output sampled data may be high:
https://www.speex.org/docs/manual/speex-manual/node9.html
Bitrate comparison:
https://www.speex.org/comparison/
Samples:
https://www.speex.org/samples/
Maybe higher compression can be achieved with better prediction, aka machine learning.