You might think it's difficult to distinguish between real human voices and AI-generated voices, and in fact, most people struggle with this. Interestingly, however, your brain may have already begun to recognize this difference, even if you're not consciously aware of it.
In a study conducted by Tianjin University and the Chinese University of Hong Kong, scientists tested the AI's ability to recognize speech in 30 participants. The results showed that humans consistently make mistakes when trying to distinguish between real voices and synthesized voices.
Even after a short training session aimed at improving their cognitive abilities, participants still struggled to make accurate decisions.
However, when scientists analyzed the neural data obtained from the electroencephalogram (EEG), they discovered something interesting: the brain was still silently processing the difference between the real voice and the AI voice.
Your brain hears things you don't realize you hear.
The study was published in the scientific journal eNeuro. In the experiment, participants listened to sentences spoken in real human voices and two different types of AI voices.
One type is a basic synthesized voice, while the other is finely tuned to sound more natural, closer to a human voice.
Participants had to press a button to guess whether each voice was real or fake, and they were wrong quite a few times. But the EEG data tracking brain activity told a different story.
After only about 12 minutes of training, the brain's neural response begins to change. The brain starts classifying the synthesized voice differently from the real voice, and this occurs at three distinct points in time: approximately 55 milliseconds, 210 milliseconds, and 455 milliseconds after hearing the sound.
These are all very early stages of sound processing, occurring even before consciousness has a chance to engage in the process of perception.
Why do your ears 'detect' it, but your brain hasn't had time to decide?
This phenomenon stems from the gap between the perception process and the decision-making process.
The human auditory system can recognize very subtle 'sound traces' in AI-generated speech. However, these signals are not yet connected to a conscious decision, that is, the 'this is a fake voice' button in our minds.
Acoustic analysis in the study revealed that real speech and AI speech differ in the frequency modulation range from 5.4 to 11.7 Hz. This frequency range is related to how the brain tracks fine details in speech, such as phonemes and the starting point of each syllable.
Even AI voices that sound very natural can't perfectly reproduce these microscopic variations. At least not at the moment.
The fight against deepfake scams
The research findings offer some positive news: humans are not entirely powerless against deepfake voice scams.
Our biological machinery – our auditory system and brain – is still capable of detecting differences. The problem is we haven't yet learned how to consciously harness this ability.
In the future, training tools could help people recognize signals that the brain has already detected, instead of just offering general advice like 'beware of deepfakes'.
In other words, the data is there, and the signals of recognition already exist. The next step is to connect those signals to the human decision-making process.
Currently, the study's conclusion offers a rather intriguing feeling: your brain is working harder than you think, and it has begun adapting to AI voices, even though your conscious mind hasn't yet caught up.