Google's Lyria model turns text into pop songs with artist vocals

Generative music has usually sucked at two critical things: keeping a song coherent across verses, and making vocals sound like real humans instead of glitchy robots.

Google DeepMind and YouTube just introduced Lyria, their most advanced music generation model to date, alongside two new experiments designed to test how AI can fit into the real music business.

Alongside the model, Google announced Dream Track for YouTube Shorts, testing AI-generated songs featuring authorized vocals from major pop artists, as well as a suite of Music AI tools for professional producers.

30 seconds of Charlie Puth with a single prompt

Music is vastly harder to generate than speech. Every second layers complex rhythms, melodies, harmonies, and multiple instruments. Keeping continuity across an entire verse or bridge has historically tripped up AI models.

DeepMind claims Lyria solves this by excelling at generating high-quality music with both instrumentals and vocals simultaneously, giving users fine-grained control over nuance and style.

In practice, YouTube is testing this directly inside Shorts with Dream Track. A select group of US creators can simply type a text prompt, pick an artist from a carousel, and get an original 30-second soundtrack featuring an AI-generated clone of that artist's voice, lyrics, and backing track.

Nine world-renowned artists signed up for the experiment: Charlie Puth, T-Pain, John Legend, Sia, Charli XCX, Demi Lovato, Troye Sivan, Alec Benjamin, and Papoose.

Charlie Puth noted that YouTube understands the need to develop technology responsibly to accelerate creativity rather than replace it. T-Pain, a longtime tech adopter, called joining hands with Google and YouTube a great first step.

Beyond Shorts snippets, DeepMind is also working with YouTube's Music AI Incubator on pro tools: humming a tune to generate a horn line, turning simple MIDI keyboard chords into a choir, or re-styling a pop track with reggaeton vibes.

Inaudible watermarking with SynthID

Whenever generative audio gets this good, deepfakes become an immediate problem. Google is deploying an audio version of its SynthID watermarking system on everything Lyria produces.

SynthID embeds an inaudible digital watermark directly into the audio waveform by converting the wave into a two-dimensional spectrogram visualization.

Importantly, Google says the watermark remains detectable even after the audio is compressed into an MP3, altered with background noise, or sped up and slowed down. It can even pinpoint whether just a specific section of a song was AI-generated.

Why it matters

Google’s smartest move here isn't the underlying model—it is the strategy around copyright and industry partnerships.

While the music industry spends millions suing unauthorized AI clone apps, Google is building a walled garden where artists explicitly opt in and help set the rules. By limiting Dream Track to 30-second Shorts clips, they sidestep the full-length track streaming debate while giving creators an instant, viral audio tool.

For creators and producers, the value is obvious: rapid background audio generation and zero-friction sketching tools that turn hums into instrumental parts.

However, we believe this curated approach comes with clear trade-offs. It keeps AI music tightly bottled within YouTube's ecosystem, and the question of long-term licensing revenue for artists whose styles train these tools remains largely unanswered. It is a brilliant defensive play, but the real test will come when these tools expand beyond hand-picked pop stars.