benchmark1a.php

<?php

function a(array $a) {}

for ($i = 1; $i < 1000000; $i++) @a('');

benchmark1b.php

<?php

function a($a)
{
    if (is_array($a)) {}
}

for ($i = 1; $i < 1000000; $i++) a('');

benchmark2a.php

<?php

function a(array $a) {}

for ($i = 1; $i < 1000000; $i++) a(array());

benchmark2b.php

<?php

function a($a)
{
    if (is_array($a)) {}
}

for ($i = 1; $i < 1000000; $i++) a(array());

gistfile5.txt

tobias@laptop:~$ time php benchmark1a.php 
real	0m6.895s
user	0m4.480s
sys	0m1.280s

tobias@laptop:~$ time php benchmark1b.php 
real	0m5.837s
user	0m3.316s
sys	0m2.444s

tobias@laptop:~$ time php benchmark2a.php 
real	0m3.447s
user	0m2.044s
sys	0m1.144s

tobias@laptop:~$ time php benchmark2b.php 
real	0m5.882s
user	0m3.220s
sys	0m2.648s

Hm. So I did a bit more testing:
with an empty arary instead of 'a':
$ time php 1.php # array
real 0m1.531s
$ time php 1.php # is_array
real 0m1.059s

However, with array('foo')

$ time php 1.php #array
real 0m0.932s
$ time php 1.php #is_array
real 0m1.205s

I tried without the @ error suppression too. That put a serious dent in the performance. Without it you're definitely right on signature checking being faster.