nds are stored in IMA 4:1 compressed format.  For MacOS, this is
supported directly in Sound Manager 3.1 and later.  For Win95/NT, QuickTime
version 2 or later should be able to handle it.

Stereo is supported for startup screen sounds and the like, but probably not
for in-game sounds coming from specific locations.

Each sound has a number of variations, one of which is picked randomly when
the sound is played.  The largest number seen in the original files is 5
(echoes of Marathon here), but the format allows for much larger numbers.

A sound tag has the following structure:

    Header (64 bytes)
    Variation names (32 bytes for each variation)
    Variation descriptions (32 bytes for each variation)
    Compressed sound data

The header looks like this:

    4 bytes     probably flags
    14 bytes    unknown
    2 bytes     the line number (starting at 0) in the corresponding
                subtitle tag (-2 if no subtitle)
    4 bytes     offset (from the start of the header) to the variation
                descriptions
    4 bytes     size of the variation descriptions and the sound data
    4 bytes     the id of the corresponding subtitle tag
    6 bytes     unknown
    2 bytes     the number of variations
    4 bytes     offset (from the start of the header) to the variation
                names
    4 bytes     size of the variation names
    16 bytes    unknown

Each variation name looks like this:

    4 bytes     possibly a 16.16 fixed point number controlling (in some
                indirect manner) the chance of this variation being chosen 
    2 bytes     unknown
    26 bytes    the name

Each variation description looks like this:

    4 bytes     unknown
    2 bytes     always 16, probably the (uncompressed) sample size
    2 bytes     unknown
    2 bytes     the number of channels (1=mono, 2=stereo)
    2 bytes     unknown
    4 bytes     sample rate as a 16.16 unsigned fixed point number
                (56220000 hex means 22050.0)
    4 bytes     the number of frames of compressed data
    12 bytes    unknown

The compressed data for the first variation starts immediately after the
variation descriptions; to get the offset for any subsequent variation, you
have to add upp the sizes of the preceding variations.  Each variation uses
34*(frame count)*(channel count) bytes of compressed data.