Artificial intelligence is not creating false photos and videos, it can also make false voices.
So far, these voices have been remarkably forced and robotic, but researchers at the AI Dessa startup have created what is by far the most compelling voice clone we've heard, perfectly mimicking the sound of MMA- commentator-turned-Podcaster Joe Rogan.
Listen to the AI Rogan clips from Dessa below, or take a test on the company's site to see if it can detect the difference between real Rogan and fake Rogan. (It's surprisingly difficult!)
In terms of making a convincing fake, Dessa chose her goal well. Rogan is probably the most popular podcaster in the world, and has recorded more than 1,300 episodes of The Joe Rogan Experience to date. That provides a large amount of training data for any artificial intelligence system.
It does not hurt that the company's engineers are obviously familiar with Rogan's favorite conversation topics. To speculate on whether or not we are living in a computer simulation, or to admire the strength of the upper body of the chimpanzees, that is all Rogan's primordial material.
But, of course, being able to convincingly falsify someone's voice also has disturbing implications. As the Dessa engineers point out in a blog post, cases of malicious use of fake voices include spam calls that impersonate loved ones; use false voices to intimidate or harass people; and creating misinformation through false recordings of politicians.
"Clearly, the social implications for technologies such as speech synthesis are massive," writes Dessa. "And the implications will affect everyone: poor consumers and wealthy consumers, businesses and governments."
The company notes that there are also benefits. These include the creation of more realistic AI assistants, faster and more accurate dubbing for TV and movies, and the design of realistic and personalized synthetic voices for people with speech disabilities.
We have contacted Dessa to learn more about its However, the company says that due to the possibility of malicious uses it will not launch its research in its entirety nor will it make its AI models publicly accessible (a position we have seen in larger artificial intelligence laboratories such as OpenAI, which controversially denied the final version of his system of artificial intelligence generation of text).
Although it can be To argue that fears about deep mistakes are exaggerated (technology has been available for years, but a forgery has not yet impacted on general policy, it is also clear that technology will only improve and will be more accessible in the future .
"At this moment, technical experience, ingenuity, computing power and data are required for models like RealTalk to perform well," says the company. "But in the next few years (or even sooner), we'll see that the technology advances to the point where it only takes a few seconds of audio to create a realistic replica of the voice of anyone on the planet."
Listen to AI Joe Rogan talk about chimpanzees tearing your balls is, strangely, just the beginning.