Or you could have an extremely long audio file so the repeat situation doesn’t occur.
The decompressed data is always the same, but the data in the dictionary used is where you store your sneaky bits.
Sure, that's still mildly suspicious. But way less than the actual music data changing all the time.
You can use a channel with lots of noise, because you can use error correcting codes to to restore the intended message.
(To elaborate with an example: sometimes a package might already drop randomly, or timings might be slightly delayed anyway.)
I guess you could also build an auto-auto-tune or auto-remix solution so the songs always have justifiable variation without needing fully generated music.
-edited because it didn't make sense before