As the development of generative artificial intelligence (AI) makes it possible to synthesize voices, ‘voice watermarking’ technology has emerged to identify them.
Tech Crunch, a professional media outlet, reported on the 1st (local time) that a voice watermarking technology produced by Resemble AI, ‘PerTH’, was proposed as a way to solve voice fraud concerns.
AI-generated voice technology raises concerns that it can be misconstrued as being actually spoken by politicians or celebrities, and misused for voice phishing or fraud.
Watermarking is the technology of printing a recognizable pattern to identify the source of an image or video. Obvious watermarks such as logos and images are common, but there are also many watermarks that are invisible to humans. It hides the pattern at the pixel-by-pixel level so that the human eye looks unmodified, but the computer can see it.
Likewise, in audio, there is a way to encode sounds that are inaudible to the human ear. However, this inaudible watermark encoding method has the disadvantage that the voice watermark may disappear even if the media is slightly modified.
For this reason, Resemble AI solved this problem by using PerTH.
PerTH is a machine learning module that analyzes the sound wave of the voice and automatically distinguishes pitches lower than the sound wave.
Watermarks are not only difficult to distinguish from real sounds, but also difficult to remove because they are closely associated with similar band voice information. On the other hand, watermarks are easily recognized with PerTH.
Resemble AI is also an AI startup that provides generated voice services that replace human voices in dubbing and audiobook services. Therefore, they said that they have created such a device to prevent their services from being misused.
Resemble AI announced that it will soon provide a PerTH watermark service to all customers.
Correspondent Juyoung Lee juyoung09@aitimes.com
