Google is testing new generative AI features for YouTube that will allow people to create music tracks with just a text prompt or a simple hummed melody. The first, Dream Track, which is already available to some developers on the platform, is designed to automatically generate short 30-second music tracks in the style of famous artists. The feature can mimic nine different artists who chose to collaborate with YouTube in its development. YouTube is also introducing new tools that can be used to generate music tracks from a hum.
According to YouTube, the Dream Track feature is currently being tested with a “small group of select US creators” and can produce tracks in the style of nine artists; Alec Benjamin, Charlie Puth, Charli XCX, Demi Lovato, John Legend, Papoose, Sia, T-Pain and Troye Sivan. In two video demonstrations, we’re shown how the prompt “A ballad about how opposites attract, upbeat acoustics” can be used to create a Charlie Puth-style track, or how “A Sunny Morning in Florida, R&B” can be used to make one T-Pain song. The software can generate lyrics, a backing track and an AI-generated voice in the artist’s style.
The idea is to use these tracks with YouTube’s TikTok-style Shorts service. In September, YouTube also announced a new AI feature called Dream Screen, which can be used to generate videos and photos as backgrounds.
A sample user interface for YouTube’s Music AI tools. Image: Google
YouTube says participants in its Music AI incubator will be able to test these tools later this year.
These new AI tools are based on a music generation model called Lyria from Google’s DeepMind. In an accompanying blog post from DeepMind, the Google subsidiary says that tracks created with Lyria carry a SynthID watermark, which is inaudible to the naked ear and can be preserved if a track is changed. So even if someone adds more noise to a track, compresses it into an MP3 file, or speeds it up, in theory it should still be possible to detect that it contains Lyria AI-generated audio.