top of page

Google Deepmind's Lyria Uses SynthID for AI Music Watermarks

On November 16, 2023, Google DeepMind announced a new AI music generation model called Lyria. It will join a set of tools released by Google this year, including the notable generative text-to-music app, MusicLM.

Google deepmind lyria dashboard
Lyria Dashboard | Credit: Google Deepmind Blog

Lyria creates instrumental and vocal arrangements, but aims for better continuity across the sections of a song. The new music it creates will be run through SynthID and encoded with an inaudible watermark, to help mark it as AI-generated audio.

Google says they want to serve songwriters and music producers. People can hum melodies into Lyria and turn them into full song arrangements. AI-generated music could become a quick way to audition new ideas during the creative process. Watch a demo of their hum-to-song feature below:

From the demos we’ve heard, it’s still not quite producing high-quality music. The songs could be pulled into a DAW and used as backing tracks, or re-created with MIDI and new sound design. But musicians don't need to start worrying... yet.

YouTube has said people in their Music AI incubator will soon be able to test the tools, probably before the end of year. The rest of us will have to wait for access to their new AI music generation tools.

Why is Google adding the SynthID watermark to their audio?

SynthID is an inaudible watermark that helps Google identify and track generative AI audio without impacting the listening experience. It was first used for AI images created by the Imagen app on Google Cloud’s Vertex AI.

Google’s text-to-music app MusicLM didn’t have watermarks when it came out in May 2023. Mounting legal pressures from the music industry are likely spurring DeepMind to innovate and create attribution IDs.

On October 18, Universal Music Group (UMG) filed a lawsuit against Anthropic, a major artificial intelligence company. The lawsuit states that Anthropic’s chatbot Claude 2 was quoting copyrighted song lyrics without a license.

Google has $300 million invested in Anthropic — Amazon has $4 billion. The ruling in this case will set a precedent for copyright law and how big tech proceeds with its operations.

Google DeepMind faced other legal challenges in 2016 after it mishandled the personal data of British healthcare patients.

On October 30th 2023, president Biden issued an executive order that called for watermarking AI generated content to help regulate the creation and distribution of “deepfakes”.

With SynthID in place, the DeepMind team seems prepared to continue exploring and creating new music ai tools. The Lyria model is likely one of several AI experiments to come.

YouTube Shorts: Create “Dream Tracks” with Lyria AI

Google DeepMind’s team partnered with YouTube Shorts to create Dream Track. This AI experiment gives fans access to artist voices through Lyria’s AI-powered music creation.

As the demo above shows, all you have to do is choose a topic and a style of music. Dream Track uses Lyria to generate 30-second soundtracks in the style of an artist from their limited roster.

Nine artists have given their permission for the tool to replicate their styles, including Alec Benjamin, Charlie Puth, Charli XCX, Demi Lovato, John Legend, Papoose, Sia, T-Pain, and Troye Sivan.

However, after reading several press releases, it’s still unclear where the instrumental music generation layer is coming from. The most likely candidate is MusicLM, or a fork of that project.

See our article on Google’s AI music datasets for an overview of how they trained their text-to-music app MusicLM, May 2023. We've also written about AI dream songs in the context of Meta's model, MusicGen.

Google outshining OpenAI, Microsoft, Adobe

Google isn’t alone in their race to create impressive new AI models.

During November 2023, OpenAI released a slew of new ChatGPT features including text-to-voice, Dalle-3, and image interpretation. Generative music was one of the few features missing from the update, actually. What happened to OpenAI's MuseNet and JukeBox teams?

Microsoft’s AI music research branch in Asia announced some small breakthroughs in AI text-to-midi generation but cannot really be considered competitive with Google and Youtube.

In November 2023, Adobe announceds a new text-to-music AI model called Music ControlNet. We reviewed Music ControlNet, but there is currently no way to test it.

It seems like Google could take the lead on text-to-music in the near future. Let us know if you agree in the comments section below.

bottom of page