You could show several faces and ask "how many are male/female?"

Another idea along these lines would be to present a scene and ask the user how many objects are in the scene. It's a very difficult problem for computers to descern individual objects in a scene, especially if they overlap.

It's also not so good for the visually impaired because you can't really provide an audio version of the challenge, whereas identifying the letters in a random string is easier to present as audio.

That's true. Maybe you could do something with profiles from FB. Show a small collections of faces one of which is your friend on FB. Then you have to select your friend. For the visually impaired you could read peoples' names.

A real person would instantly recognize one of their friends and even the guy getting paid to solve captchas would have a hard time.

Of course, you then have the problem of getting someone to connect FB to the CAPTCHA service for the first time instead of abandoning the form.

Do captcha bots ever use the audio string? Could you not hook up a voice recognizer to the audio portion?