Finding minute differences between pictures is a much harder problem than matching voices to those you've heard before. In the second one there are lots of possible heuristics you could use, and the attacker would have to prevent all the most common ones at least.
They're not even minute differences. One example toggles between two pictures of an airplane, with the second having the engines photoshopped out. Just a one-second delay is enough to make it difficult to notice what changed.
Now of course it's easy if the two pictures (or voices) are completely different. But who would be suspicious if the two voices sounded similar?
