Amazon demoed an Alexa feature at re:MARS in which a child’s late grandma comes out of the speaker to read a tale. The system can learn to imitate someone’s voice from just one minute of recorded audio, because Amazon treats the problem as voice conversion rather than speech generation. That matches Amazon’s ICASSP 2022 paper on Voice Filter, which adapts an existing text-to-speech voice to a new speaker with as little as one minute of their speech. Amazon did not say when the feature ships or how consent would work for people who can no longer give it. Rachel Tobac of SocialProof Security warned that it will likely be used for impersonation, and the easier it is to use, the more it will be abused. Your boss calling to ask for a password just got a lot more convincing.
