Lyria

Lyria is a computer program from Google that generates music. You type in a description, for example “calm jazz with saxophone and rain in the background.” A few seconds later, a piece of music comes out that never existed before. It was developed at DeepMind, Google’s AI division. Lyria doesn’t just produce a melody as notes, but directly generates finished sound complete with drums, bass, instruments, and in some versions, vocals as well.

What Lyria means for the music industry

Music was long an area where machines performed poorly. Programs had already been able to produce passable text and images for some time. Music is harder because it has to hold together over the course of minutes. A song can’t suddenly change key in the third chorus. It is precisely this kind of long memory that models like Lyria have now managed to master reasonably well.

Economically, this affects one particular market above all: background music. Millions of videos on YouTube, TikTok, or in advertisements need music that nobody really notices. Until now, people bought licenses for this from providers of so-called stock music. An AI can deliver such music in seconds, at almost no cost. For composers who make their living from exactly these kinds of jobs, this is a serious threat.

At the same time, the legal situation remains unresolved. Such models were trained on huge amounts of existing music. Whether record labels need to be paid for this is the subject of several lawsuits. Google has therefore signed agreements with labels and individual artists early on, rather than simply proceeding without them. This sets Lyria apart from some competing products that launched without such arrangements.

From training material to finished audio track

Lyria has learned from a very large number of music recordings which sounds typically follow one another. This was never about memorizing entire songs. The model stores patterns: what reggae sounds like, what a chorus sounds like, what a transition between two parts sounds like. When generating something new, it assembles it from these patterns.

Technically, the model doesn’t work directly with the raw audio track. One second of music consists of tens of thousands of individual data points, which would be far too much computational work. Instead, music is translated into more compact building blocks, a kind of shorthand for sound. The model predicts the next building block step by step. In the end, a second program translates these building blocks back into audible sound.

Every generated file also receives an inaudible watermark called SynthID. You can’t hear it, but a detection program can recognize it again, even after compression or editing. This makes it possible to later prove that a piece originated from a machine. This is important, because otherwise no one would be able to distinguish real music from generated music anymore.

Lyria in Google’s products and in the headlines

You encounter Lyria most directly in MusicFX, a free tool in Google’s experimental lab. There, you enter a description and receive short pieces in return. A variant called Lyria RealTime generates music continuously while you adjust dials for mood and instruments. This resembles an instrument more than an order form.

The technology became well-known through Dream Track, an experiment on YouTube Shorts. There, selected users could generate short soundtracks that sounded like the voices of well-known musicians. Nine artists had explicitly agreed to this, including Charlie Puth and T-Pain. The project remained small, but showed where things are headed.

In business news, Lyria usually comes up in comparison with Suno and Udio, two start-ups with similar products. Both were sued by major record labels. A common misconception here: Lyria is not a product you can buy, but a model working in the background. Users always only see the application built on top of it, much like an engine that’s installed in different cars.