People are poorly equipped to detect AI-powered voice clones.

Area of Science:

• Speech processing
• Human-computer interaction
• Artificial intelligence

Background:

• Generative artificial intelligence (AI) is rapidly advancing, producing increasingly human-like content across various modalities.
• The realism of AI-generated audio, particularly voices, is a critical area of ongoing research and development.

Purpose of the Study:

• To evaluate the perceptual realism of AI-generated voices.
• To assess the ability of human participants to identify AI-generated voices and match them to their original speakers.

Main Methods:

• Conducted a series of perceptual studies involving human participants.
• Presented participants with recordings of both AI-generated and human-generated voices.
• Collected data on voice identity matching and the accuracy of AI voice identification.

Main Results:

• Human participants demonstrated a limited ability to consistently identify AI-generated voices.
• Participants perceived the identity of AI-generated voices to be the same as their real counterparts a significant portion of the time.
• Accurate identification of a voice as AI-generated occurred only about [Formula: see text] of the time.

Conclusions:

• AI-generated voices exhibit a high degree of realism, challenging human perception.
• The indistinguishability of AI voices has significant implications for authenticity and trust in digital communication.
• Further research is needed to understand the nuances of AI voice perception and develop robust detection methods.