Back in April, a new generation of speech synthesis technology called Lyrebird threatened to render voice passwords useless. It was developed at the Montreal Institute for Learning Algorithms (MILA) lab at the University of Montreal and claimed to perfectly copy any voice given one minute of spoken audio.

Lyrebird wasn't actually perfect, but it was very good and is sure to keep getting better. The University of Washington just unveiled a new AI system that's just as good as Lyrebird, but focused on lip-syncing audio to recorded video.

The system developed at the university can successfully generate a photorealistic video of a person talking using a pre-recorded audio file. It's good enough that the video is lip-synced perfectly and it would be hard to tell it wasn't real unless you were looking for it.

Here's an example using former President Barack Obama:

In order for this to work, the research team trained a neural network by allowing it to watch many hours of Obama's weekly address footage. It was then able to take an audio file of his speech and map it on to another video matching the mouth shapes it learned Obama makes when he talks. The result, as you can see, is impressive.

On its own, the system can take audio already spoken by a person and map it on to other videos of them talking. But what about if you combined this with Lyrebrid? If the system is already trained on Obama's mouth movements, and Lyrebird can copy his voice, you could get the former President to say anything and for it to look real.

While that's a scary thought, there are much more practical uses for this lip-synching system. For example, it could be used to improve streamed video over poor connections where the audio is prioritized and the video suffers. As long as the system is trained on the person's movements, it could enhance the video to look high quality and the viewers wouldn't know it wasn't real.