This issue was first reported by Karl T in:
http://lists.w3.org/Archives/Public/public-audio/2013JulSep/0568.html

[[
When I first read that the DelayNode "delays the incoming audio
signal by a certain amount", I imagined something like

(1)     y(t + d(t)) = x(t)

for input x(t), output y(t) and delay d(t).

However, I get the impression that delay lines are often
implemented as

(2)     y(t) = x(t - d(t)).

So I think I may have been reading more into the statement above
than was intended.
]]

Ehsan confirmed in:
http://lists.w3.org/Archives/Public/public-audio/2013JulSep/0592.html
that the latter is implemented. 

The spec prose should be clarified to reflect this.