On Mon, Feb 11, 2013 at 11:35:41AM -0800, Aaron L. wrote: It does both in a sense. Not on the original samples that make up a waveform, but on a spectral representation of the signal. What happens (roughly) is that the signal is chopped up into small fragments, and for each one the spectrum (fourier transform) is computed. Transmitting or storing the spectra instead of the signal would take exactly the same amount of information. But the spectrum can be 'simplified' without affecting audio quality quite easily. Depending on the relative levels, some frequencies will 'hide' others - you don't hear the weak ones. So those can be deleted from the spectrum ('dropping samples'). Also it doesn't matter much if some frequency has a level of e.g. -35.4 or -35.3 dB, so these numbers require much less bits than a sample would (rounding). The result is that the simplified spectrum needs much less data than the original signal. The decoder just recomputes the signal from the spectrum data. Ciao, -- FA A world of exhaustive, reliable metadata would be an utopia. It's also a pipe-dream, founded on self-delusion, nerd hubris and hysterically inflated market opportunities. (Cory Doctorow)