It may be possible to verify the use of Google’s AI technology in creating a recording by employing SynthID to imprint the voice generated by DeepMind’s Lyria model.
Journalist ByJon Porter, with a background spanning five years in reporting on mechanical keyboards, online platforms, EU tech regulations, and consumer technology launches, sheds light on the implementation of SynthID to distinguish AI-generated content in retrospect.
The utilization of Google DeepMind’s AI model, Lyria, to produce audio, including tracks generated through YouTube’s novel audio creation functionalities, will now incorporate a SynthID watermark. This watermark is designed to remain identifiable even amidst compression, alterations in speed, introduction of noise, as outlined in a DeepMind blog post. Notably, the watermark is imperceptible to the human ear, preserving the original listening experience.
In the realm of AI, tools like SynthID and other watermarking solutions play a critical role in mitigating the pitfalls associated with generative AI technologies. For example, President Joe Biden’s executive directive on AI advocates for the establishment of comprehensive guidelines for watermarking AI-generated content by governmental bodies. Despite the allure of this field, contemporary technology still grapples with imperfections in safeguarding against fraudulent activities.
DeepMind’s description of SynthID’s audio application involves transforming sound waves into a visual representation depicting the evolution of frequency ranges over time. This innovative approach, according to DeepMind, sets it apart from existing methodologies in the domain.
Following the alpha release of SynthID for images created using Imagen on Google Cloud’s Vertex AI, Google revealed the integration of hashing capabilities into AI-generated audio. While DeepMind acknowledges that this method may not be impervious to extreme alterations in images, the watermark remains resilient against enhancements like cropping or resizing.
