[ad_1]
Three seconds: this is the time it takes for this AI to imitate any human voice.
Microsoft has just presented a speech synthesis model capable of faithfully simulating a person’s voice from a very short audio sample.
This AI model, called “VALL-E”, has been trained with more than 60,000 hours of conversation in English, “hundreds of times more than existing systems”, describes Microsoft.
VALL-E not only mimics the intonation of a voice, it is also able to “preserve the speaker’s emotion and the acoustic environment of the acoustic prompt in the synthesis”.
Technology is not ready for the real world
This language model enables various speech synthesis applications, such as speech editing and audio content creation, by combining other generative AI models like GPT-3.
However, Microsoft does not leave ethical considerations aside. “Since VALL-E can synthesize speech retaining the speaker’s identity, it may carry potential risks of misuse of the model, such as voice identification spoofing or speaker impersonation. a specific speaker,” Microsoft warns.
At this stage, the AI is not publicly available. To be used in the real world, Microsoft indicates that this system “should include a protocol to ensure that the speaker approves the use of his voice and a model of detection of the synthesized speech”, in order to be able to tell quickly if what you hear is a human voice or a synthetic voice.
Cyber risks
Microsoft’s AI can cause security issues if it gets into the wrong hands. Imposters could, for example, associate VALL-E with deepfake technology to create fake videos. “Deepfakes,” the process of artificially recreating audio, visual, and video content of human beings through AI, have been a potential identity theft threat for several years.
VALL-E could also allow cybercriminals to imitate the voice of a potential victim on the phone, in particular to authorize operations.
Recently, a team of cybersecurity researchers discovered that openAI’s chatGPT was also starting to be hijacked to write malware.
[ad_2]
Source link -97