Specifically, I was watching “The Simpsons” earlier and wondering if, for example, they recorded Hank Azaria while he was doing the voices of Moe Szyslak, Professor Frink, and Chief Wiggum, would you be able to identify him via voiceprint even though he’s drastically modifying his voice in different ways for the different characters?
All I know about voiceprint analysis, I got from watching “Law & Order”, and it doesn’t address the whole “guy talking in cartoon character voice” issue, so I thought I’d bring it up here.
I’d understand that you can. From what I know about voiceprint, it’s like fingerprint: The computer takes the waveform of your voice and compares it, point by point, to a known good waveform. If enough of the points match, the lock is disengaged (logically or physically).
I can’t imagine Moe Szyslak and Clancy Wiggum being able to unlock each other’s voiceprint ID systems.
I dunno about that, David Simmons. Maybe so for you or me, but a trained voice actor? I remember reading somewhere once that James Doohan, who played “Scotty” on Star Trek, could keep up his fake Scottish accent all day if necessary, never once breaking out of it into his normal voice.
I would guess that you could. Even being drunk or much much older is enough to throw off voiceprint analysis.
It must be true! I learned it on the teevee!!!
Seriously, I did see something last week about voiceprint analysis being a good tool for comparing a known voice with a recording that was within a few years of the known voice, but not so great when you have a lot of variables like illness, age, and being stoned in the mix.
This may be completely inaccurate, but in The Dark Half they talk about voice-prints from the same time frame as being identical, even though the voices sound nothing alike. I recognize the dangers of getting one’s science from horror novels, but Steve’s usually pretty good about not just pulling stuff out of his ass.
There are several problems with speaker verification systems, but one of the objectives is to make it not succeptable to imitators. That is, if you are speaking in a funny voice it may not recognize you as you, but you wouldn’t want it to accept someone doing an imitation of you (or playing a recording of you) as the real deal. This is a very hard problem, but any system that didn’t address it would be of limited usefulness for security.
I don’t believe that recordings are compared “point-by-point”, Derleth. There is a lot of variation in your own voice even throughout the day (especially when you’re ill), and this wouldn’t be very robust to those variations. Essentially, they model the activity in the vocal tract when you are speaking a known word (like your name). The idea would be that each person’s vocal tract produces speech sounds with sllightly different spectral characteristics than everyone else. Even if your voice was lower than normal (say, when you first wake up), the spectral characteristics are largely the same, just translated down a little. If it were someone else trying to imitate you, their vocal tract would be physically shaped a little differently, which would produce slightly different spectral characteristics (even if they were imitating your accent and the cadence of your speech).
Techniques like linear predictive coding and hidden markov models are common in speech analysis if you want to look into details. As far as I know, no one has shown that we are as identifiable by our voice as we are by our fingerprints or our DNA. I don’t know that I believe that this will ever work well, but this is my understanding of how it tries to work against imitators. I googled “speaker verification algorithms” and it came up with a bunch of stuff if you want to learn more.
Oh sure, and an expert locksmith can get into your car in 30 seconds, and on and on with specialists. If you know that the suspect is a trained actor you have to use other methods.
I heard Rich Little describing a test he did with a voice identifier when he imitated Reagan, or maybe it was Nixon, and his imitation was compared with a high quality recording of the real voice. Not even close, and everyone agreed it was an excellent sounding imitation.
I don’t think it is worth the expense and years of study to become a trained actor in order to fool a voice identifier once.