Imitate a voice with a vocoder

August 9, 2005 9:27 AM

Could a vocoder be used to imitate some elses voice? If you had a recording of someone speaking, could you make a sample of their "carrier" sound and modulate it with your own voice? You would have to vary the pitch of the carrier to sound like normal speech, but assuming you could do that (a big if, I know), it would be fairly convicing no?
posted by phrontist to Media & Arts (5 answers total)

I'd say not. I'm no professor of linguistics or phonetics, though I have an interest in each, but there's a lot more to speech than timbre, which is all you could hope to replicate.

Of most importance would be the different way that we pronounce different phonemes. This can often vary between family members, let alone between the population at large. Your pronunciation of all the phonemes is almost like a vocal fingerprint.

If you were a skilled impersonator you would definitely produce a much better sound than what you're suggesting, I'm afraid. In fact, I'd hazard a guess that even a bad impersonation would be better.
posted by wackybrit at 9:41 AM on August 9, 2005

Vocoders are rather crude devices and have a limited number of frequency bands. They can't shape the subtle overtones that enable us to recognize various speakers.

However, there are devices (and software plug-ins) that, using physical modeling of the human vocal tract, strip the resonances from a voice and apply new resonances characteristic of someone else's vocal tract. The math involved is the same as is used in "harmony processors" like the DigiTech Vocalist except in the Vocalist's case, you're changing the pitch but keeping the original singer's resonances. (Although actually I believe many such devices can work both ways.) You'd have a hard time making it sound like a specific person's voice, but you could probably make it sound like a different person pretty easily. The Vocalist is old and discontinued, but there are lots of ways to do the trick today.
posted by kindall at 9:50 AM on August 9, 2005

a vocoded bit of audio at this point always has many artifacts. theoretically a phase vocoder (like the one used in Soundhack and Native Instruments Vokador) should be able to do something that sounds organic (although most likely not like what you describe) but thus far all phased vocoding algorithms use 1024 or fewer analysis bands for the FFT processing they do and as a result they sound artifacty and glitchy.
posted by n9 at 9:59 AM on August 9, 2005

You might be able to use the PRAAT package, which was recommended on the homepage of one of our own resident geniuses, effugas.
posted by sonofsamiam at 10:17 AM on August 9, 2005

Best answer: Check out this page. It's pretty impressive. They take a sample recording of the target voice, then the source voice says the same phrase as the target. The two recordings are used to generate a "diff" between the two, which is then applied to make any arbirtrary statement sound like the target voice.

I actually tried to contact the author without success, the email address is old. I was interested in seeing the code behind it, or at least a demo I could play with.
posted by knave at 10:22 AM on August 9, 2005 [1 favorite]