Google Launches Lyria 3.5 in Flow Music

Google has launched Lyria 3.5, its newest music generation model, rolling out today in Google Flow Music with what the company describes as significant advancements across musicality, lyrics, and vocal quality.

The release is Google's latest move in AI-generated music, a category where the quality gap between synthetic and human-recorded tracks has narrowed rapidly. For Google, Lyria 3.5 is a statement about where it believes the frontier sits in each of the three pillars of song generation: instrumentation, words, and the human voice.

What Lyria 3.5 changes

According to Google's announcement, the model delivers four main improvements over its predecessor:

Improved musicality. The model can now create "richer, more complex melodic structures that sound more natural," according to Google. Musicality — the coherence and naturalness of generated melody, harmony, and arrangement — has long been one of the hardest problems in AI music. Melodic output from earlier generations of music models often drifted into repetitive or structurally flat patterns. Google's claim here is that Lyria 3.5 produces melodic structures with more complexity while sounding more natural to the ear, not less.

Enhanced lyrics. Lyria 3.5 generates higher-quality lyrics with what Google calls "improved prompt adherence and structural awareness." Prompt adherence means the model follows the user's instructions about lyrical content more faithfully. Structural awareness means the lyrics conform better to the conventions of song form — verses, choruses, and bridges landing where a listener expects them — rather than producing text that scans as generic poetry set to music.

Improved vocals. The new model brings "more expression and emotion to your songs with more realistic and emotionally nuanced vocals, plus improved pronunciation," Google says. Synthetic vocals have been a persistent weak point for AI music systems. Flat delivery, mispronounced words, and emotionally neutral phrasing are common failure modes that make generated tracks easy to identify. Google's emphasis on emotional nuance and pronunciation suggests the company is targeting exactly those tells.

Creative control. Users can now more easily control the tempo and duration of generated outputs. This addresses a practical complaint about generative music tools: users often want a track that fits a specific length or pacing — a 30-second spot, a three-minute song at a particular BPM — rather than accepting whatever the model produces. Direct control over tempo and duration moves Flow Music closer to a production tool and further from a slot machine.

Where to use it

All of these capabilities are available today in Flow Music, Google's AI creation environment for music. Google frames the goal of the product plainly: "where we want to help you create songs you love, with creative control." That framing — songs you love, with creative control — signals the company's positioning of Flow Music as a tool for expressive intent rather than passive generation.

Why it matters

Music generation has become one of the most competitively active corners of the generative AI market. Voice cloning and full-song synthesis have improved fast enough that the technology now touches real commercial stakes: production music libraries, advertising, game soundtracks, and independent artists' workflows are all areas where generated audio is already in use.

Against that backdrop, the specific improvements in Lyria 3.5 map onto the three problems that determine whether AI music is usable in practice. Complex, natural-sounding musicality determines whether a track can hold a listener's attention. Prompt adherence and structural awareness in lyrics determine whether a creator can get the song they asked for on the first or second try, rather than the tenth. Realistic, emotionally nuanced vocals determine whether the output can pass as a finished recording rather than a demo. And tempo and duration control determines whether the output fits the creator's actual use case.

Google's decision to ship all four improvements in a single release, rather than iterating on one dimension at a time, suggests the company sees the bar for usable AI music as requiring simultaneous progress across melody, language, and voice.

The stakes are not only commercial. As AI-generated music becomes indistinguishable from recorded performances on more listens, questions about attribution, licensing of training material, and disclosure of synthetic content grow sharper. Improvements of the kind Google claims in Lyria 3.5 — particularly in vocal realism — are the same improvements that make those policy questions urgent.

What to watch

The immediate test of Lyria 3.5 is empirical: whether tracks generated in Flow Music actually sound richer, follow prompts better, and deliver more convincing vocals than its predecessor on real creative work. Users can try the model today in Flow Music, and the tempo and duration controls give them concrete parameters to test against their own needs.

For Google, the release is one step in an ongoing race. Each improvement in musicality, lyrics, and vocals narrows the distance between a text prompt and a track a listener cannot distinguish from studio output — and shifts the conversation from whether AI music is possible to what it gets used for.