Permission286

As R&D refines the technology and it becomes more scalable, media synthesis is morphing from a novelty into an expanding market. Companies like Amazon, Microsoft, Papercup, Deepdub, and Synthesia have created projects such as ad campaigns featuring an AI-generated Snoop Dogg and David Beckham’s voice translated into nearly a dozen languages. They’ve also partnered with news organizations including Sky News, Discovery, and Reuters to develop prototypes for automated news and sports reports.

It only takes a few seconds to minutes for AI to imitate a person’s prosody. Baidu’s latest Deep Voice service can clone a voice with just 3.7 seconds of audio samples, and WellSaid Labs, which launched as a research project at the Allen Institute for Artificial Intelligence, can create a 10-second audio file from roughly 4 seconds of speech.

To create synthetic voices and videos, companies use a combination of AI and machine learning techniques including generative adversarial networks (GANs). GANs are two-part machine learning models consisting of a generator that creates samples and a discriminator that attempts to differentiate between these samples and real-world samples. Top-performing GANs can create realistic portraits of people who don’t exist, or even snapshots of fictional apartment buildings.

Some companies have taken steps to prevent the misuse of their platforms. For example, Synthesia says it vets its customers and their scripts and requires formal consent from a person before it will synthesize their appearance, and the company refuses to touch political content. WellSaid also doesn’t create voice avatars without actors’ permission and subscribes to the “Hippocratic Oath for AI” proposed by Microsoft executives Brad Smith and Harry Shum. As for Resemble, it released an open source tool that detects deepfakes by deriving high-level representations of voice samples and predicting whether they’re real or generated.