Mp3 relies on the fact that loud sounds mask spectrally and temporally adjacent softer sounds (psychoacoustic masking) and allows the noise floor to rise for so

Mp3 relies on the fact that loud sounds mask spectrally and temporally adjacent softer sounds (psychoacoustic masking) and allows the noise floor to rise for some bands in such conditions to reduce the required number of bits to encode the band signal. Now if the masking sound disappears or changes, the crap is no longer masked. In other words the psychoacoustic principles work only when we don't alter the signal too much after the codec. Many diminutive speakers make use of techniques like multiband compression (dynamics, not entropy) to produce "larger" sound. That wreaks havoc on the psychoacoustic model of lossy codecs.