May 23, 2023
The AI Revolution and Its Double-Edged Sword
The advent of artificial intelligence (AI) has revolutionized many aspects of modern life. However, the sophistication of this technology also brings alarming potential for misuse. A recent example in the United States Senate vividly illustrates this point. During an opening statement at a Senate hearing on AI, Senator Blumenthal used a recording that sounded like him, but, as he explained after the stunt, was produced by a so-called voice cloning tool (deepfake audio) and whose content was written by ChatGPT.
This scenario, described by the senator as “eerie, even creepy,” is indicative of the pace at which technology is outstripping regulation. AI, specifically its deepfake audio applications, has evolved to the point where artificial intelligence models can craft speeches, like OpenAI’s ChatGPT or Google’s Bard, and coupled with an AI-based Text to Speech, can mimic voices with startling accuracy to the human ear.
The Growing Threat of Deepfake Audio Applications
While AI continues to deliver tremendous benefits across industries, Blumenthal’s demonstration raises crucial concerns about its potential misuse, like spreading misinformation or exploiting personal data. This has sparked discussions about the implications of this sophisticated technology and the necessity for regulatory measures to control its use.
But what are the plausible solutions to this challenge of detecting AI-generated voice versus a real one? The answer might lie in AI itself.
Benefits of Deploying ValidSoft’s AI-based Voice Verity™, Against AI Voice Mimicking
Our answer is to deal with the advance of speech AI through the use of AI. To any adversarial attack, there is a corresponding detector that can be trained to deal with the attack. This is the bread and butter for AI. Here is our recipe: expose our well-designed deep neural network to numerous examples of real and fake voices, assign our expert team of AI specialists to ensure the models are trained with the right balance of data to avoid the usual AI pitfall (bias, overfit), add years of experience in the domain and some of our proprietary speech processing techniques.
Applying advanced deepfake detection to combat the misuse of AI voice cloning technology has several advantages. First, it can detect anomalies that may be inaudible to the human ear. Even the most accurate AI voice clone will struggle to perfectly replicate the exact naturalness of a human voice. In previous studies, we have shown how the latest so-called voice cloning tools leave traces, which we call anomalous signal artifacts.