The further away from the "ideal" reference listening case of high fidelity headphones in a quiet room, the less likely that the assumptions behind the psycho-acoustic model used in the lossy codec are valid. Frequencies can be missing because of shitty speakers, room cancellation, or deliberate post-processing, and expose sounds that the codec thought would be masked. On my surround sound, the difference between a 320kbps mp3 and a FLAC is not only audible but obvious when played in Pro Logic 2 mode, which exploits phase information to place different sounds in different speakers.