Lyrebird Algorithm Can Mimic Voices After Hearing Only 60 Seconds of Audio

Canadian Startup Lyrebird has developed a new algorithm which can mimic voices with scary accuracy after sampling only 60 seconds of audio speech. The algorithm can copy the voice quality but also captures the cadence very well for only a 60 second sampling. Lyrebird also claims that it is robust enough to learn even from noisy recordings. It is also capable of pronouncing the same sentence with various intonations.

On their demo page, Lyrebird has several politicians speaking about the virtues of Lyrebird and also has various selections and examples of the algorithm in action. Barrack Obama’s simulated voice is the most accurate, perhaps because of his very notable cadence while Donald Trump’s simulated voice is a bit bumpy on some parts, especially in sustained sentences but otherwise difficult to discern from the real thing on some samples. The simulated voice of Hillary Clinton is definitely the least accurate that is why most of the demos on their page so far are Obama’s and Trump’s.

Although exciting, this kind of technology is sure to fuel the fake news paranoia even further. Compounded with other technologies such as the Face2Face project developed by Stanford which could super-impose an animated face on live video that is difficult to discern from the original, internet trolls and fake news makers would certainly have a field day.