Stereo signals need some special considerations to exploit all redundancies:

- The
*masking threshold*can be assumed to be the same for both channels. Thus, the mean of the two channels is fed to the masking threshold estimator. The threshold is of course only transmitted once for both channels. - Often, the audio is centered, i.e approximately equal in both
channels. This is used by encoding and
*d*=*l*-*m*, i.e the*mean*and*difference*signals.

Experiments has been
performed on the assumption that, in every frequency band, the signal is
only a *panned* monophonic signal, i.e the same signal is in both
channels except for a scaling factor. The left and right channel is seen
as two vectors, orthogonal to each other (see figure 7). The signal data in a band is
used to least-square-fit a vector between the two vectors, representing the
most probable direction (the optimal direction would be found with an SVD,
though). This vector, and an orthogonal counterpart,
is chosen to
be the new basis for the signal, and the left and right signals are
transformed to the new basis. Even without counting the bits for the basis
information, this did not improve the results from just the
mean-difference approach.

**Figure:** The left-right stereo model, discussed in section 4.3.

Sat Mar 7 16:27:43 PST 1998