From Hydrogenaudio Knowledgebase
Tonality estimation algorithms in psycho acoustic models try to measure the tonality of signals. If a signal is just a sine wave, it's totally tonal and it has not much masking capability.
If a signal is very noisy, it has a high masking capability and it can mask frequency content depending on the noise level and amplitude (volume) of the signal. Of course the higher masking capability there is, the lower bitrate can be used, because the content which is masked doesn't have to be encoded at all.