I think Google's voice search already does this (feeding into audio captchas), but obviously there is a smaller crowd from which to source free audio recognition assuming that people who are both sighted and hearing typically prefer visual captchas due to environmental constraints.

Most of the privacy concerns voiced in this threat don't sound like anything new to me, but this one made me take notice. You're suggesting that Google takes random queries from individuals and serves them as captchas to other random people? That sounds like a privacy disaster. Most of the time the queries would be anonymous, but it's certainly not guaranteed there wouldn't be identifying information.

Sorry, I could be completely wrong. After further research, it seems like today's audio captchas consist of distorted output from TTS engines.