> You actually can remove about 97-98% of the data from sound record and human will not notice anything at all! The catch? These 97-98% of data will depend not only on human, it depends on the state of said human. Tomorrow (when humidity or pressure will be different) you'll need another 2-3% of the data to hear it as “perfect recording”.
> This means that any sound record with “good enough” quality contains enormous amount of redundancy.
You're proving my point. If that "redundancy" can actually influence how a human would hear the sound, it isn't available for watermarking. The watermark isn't designed for a specific human in a specific state; it has to be transparent to nearly all humans, in nearly all states.
The bits used to transparently watermark an audio file cannot have a noticeable influence on the sound, by definition. Ergo, if they are not removed during compression, the compression is suboptimal.