For generations, the ability to recognize and determine the authenticity of a voice has been an important skill. However, with the rise of artificial intelligence (AI) technology, an increasingly popular form of AI-generated voices has recently set its sights on the traditional method of verification.
In an attempt to help verify the authenticity of AI-generated voices, a London-based startup called Lyrebird has developed an ‘inaudible’ watermarking technique that can accurately identify a voice as AI-generated. This watermark is embedded inside the audio files and is completely inaudible to the human ear. It works similarly to the audio fingerprinting technique used to identify sources of illegal music downloads.
The inaudible watermark consists of a small portion of frequency-modulated noise that is digitally embedded into the AI-generated voice. The algorithm that produces the watermark is tuned to the output device so the sound is only recognizable by the same device. It uses two distinct components that modify the audio signal, consisting of a “signature” component for encryption, and a “detail-specifying” component for comparison.
The signature component is used to determine whether or not a voice is AI-generated, while the detail-specifying component helps to identify the individual voice. This watermark can be used by voice authentication systems as another layer of security when authenticating identity, ensuring that the voice is from a legitimate source.
The technology is already being used by major companies such as Google, Microsoft, and Amazon for authentication purposes. It is hoped that the use of this watermark will help to make AI-generated voices more trustworthy and ensure that the authenticity of a voice is not being compromised.
While this watermarking technique may not be perfect, it is a step in the right direction to make AI-generated voices more secure. As this technology continues to improve, it might just be the key to making sure that nothing is lost in translation.
