Well, the difficult part will be getting the phone to recognize that it's voice. You can set the voice recognition system as the input, instead of the mic, which might be able to do that. I don't think so though, because (I actually read all about this yesterday) the phone doesn't actually do the recognizing, it just opens up a live stream (like a phone call) to the Google servers, and they do the recognizing.
Also, the information that I have found so far points to the conclusion that Android does not support analysis of live audio from the mic. All these other apps that seem to be "live" are actually just taking a bunch of small samples and analyzing them really quickly so that they seem live. A 500 millisecond sample every 300 milliseconds seems to be common.
Luckily, on the side of my programming job, I'm also a sound technician, so I can tell you that (if you were willing to put in the work) there is a way to detect actual voice as opposed to just sound. Every voice is split into a few distinct ratios of frequencies which all combine to make the voice we hear, and every voice's ratios remains pretty constant, while each individual voice's ratios are different (which is why voice-based passwords work). So, if you were able to take a sample, break it up into frequencies of about 10hz each, and watch for the amplitude of each, and when you got a frequency/amplitude pattern that looked similar to a voice instead of just "white noise", you'd be in business. DOING that however, doesn't seem like it'd be easy at all. Something similar has been done before with the app called SpectralView, which displays the audio spectrum all broken up.
Also, as you can see by using the Voice Search, a voice also fluctuates a lot in how loud it is. You could look for that, but it wouldn't be as reliable.
In conclusion, how do you analyze it? Well, you would have to look for a pattern in the frequencies that looks like a voice. How do you do that? Well, to be honest, I don't know for sure. Sorry.
与恶龙缠斗过久,自身亦成为恶龙；凝视深渊过久,深渊将回以凝视…
