Google Deepmind Launches Lyria Music AI Model to Advance and Breakthrough Music Creation

Google Deepmind has launched Lyria, an artificial intelligence model for music generation, aimed at enhancing the creative process for musicians and artists.

Google Deepmind has collaborated with YouTube to launch two music artificial intelligence experiments: Dream Track, a project conducted on YouTube shorts, and Music AI Tools, a set of tools for artists, songwriters, and producers.

The Lyria model aims to create high-quality music with instrumental accompaniment and vocals. According to Google Deepmind, Lyria supports various styles, ranging from heavy metal to techno music to opera. The company states that it can maintain complex rhythms, melodies, and vocals in phrases, verses, or longer paragraphs.

Google Deepmind is testing Lyria with YouTube as part of the Dream Track project. The goal of this experiment is to explore new ways for artists to create music. Users input a theme and style in the model interface, select an artist from a carousel, and compose a 30-second soundtrack for a YouTube short.

The Lyria model generates lyrics, background music, and AI-generated sounds in the selected artist's style. Participating artists include Alec Benjamin, Charlie Puth, Charli XCX, Demi Lovato, John Legend, Sia, T-Pain, Troye Sivan, and Papoose.

Google Deepmind's researchers are also collaborating with artists, songwriters, and producers from the YouTube Music Artificial Intelligence Incubator to explore how generative AI can assist in the creative process.

They are developing a series of music artificial intelligence tools that can convert audio from one music style or instrument to another, create instrumental and vocal accompaniments, and generate new music or instrument fragments from scratch. These tools aim to help artists transform their ideas into music, such as creating guitar melodies from humming.

Deepmind's SynthID tags AI audio

All content generated by Lyria is tagged with SynthID, the same technology toolkit used by Imagen on Google Cloud's Vertex AI to identify AI images. This watermark is visible to machines but invisible to humans.

Similarly, SynthID marks AI-generated audio by converting the audio waveform into a two-dimensional visual representation that shows how the sound spectrum changes over time in a way that is inaudible to human ears and does not affect the auditory experience.

Deepmind notes that the watermark should remain recognizable regardless of modifications to the audio material, such as adding noise, MP3 compression, accelerating, or decelerating the track.

SynthID can even help identify song sections generated by Lyria. Google Deepmind states that this "novel approach" is particularly rare in the audio field.

A few days ago, YouTube released new regulations regarding the handling of AI-generated audiovisual content on its platform. The video platform emphasizes the strengthening of labeling requirements and requires identification of AI content through both manual and automated means.

For YouTube, the music generator can also serve as a strategic tool to gain better control over AI-generated music on the platform, allowing users to have precise control over generated content, including tagging.