The psychoacoustic model is the part of a lossy encoder that decides what to throw away, based on how human hearing actually works. It predicts, frame by frame, which sounds the ear will and won't perceive — chiefly through masking, where a loud tone hides quieter ones near it in frequency or time, and through the threshold of hearing, below which nothing is audible. The encoder then allocates bits accordingly, coding audible detail accurately and quantizing masked detail coarsely or not at all, so the missing information is, ideally, information you could never have heard. The quality difference between two codecs at the same bitrate is largely the quality of their psychoacoustic models.
