Google launches Lyria 3.5, an AI music model with sharper vocals and structure control

The update focuses on three areas: music quality, vocal performance, and lyric accuracy. Lyria 3.5 turns text prompts into full compositions while preserving song structure, and Google says both vocals and instrumentation now sound more natural throughout a track.

Users get finer control over tempo, style, mood, and other parameters, and the model supports multiple genres and languages. According to Google, melodies are richer and more complex than in prior versions, lyrics track user prompts more closely, and vocals carry more emotional range with clearer pronunciation.

One new feature lets users submit an image as inspiration for a composition, translating a visual mood or scene into a musical direction rather than starting from text alone. Users can also define a song’s structure directly, which trades some creative unpredictability for a more expected result.

Lyria is Google’s broader family of AI music models, which has previously powered features like MusicFX and Music AI Sandbox — Lyria 3.5 extends that same lineage rather than introducing a standalone product.

Every track generated with Lyria 3.5 carries a SynthID digital watermark, Google’s system for marking AI-generated content so platforms and users can identify it after the fact. Google also says the model isn’t built to reproduce the style of specific artists, a guardrail aimed squarely at the copyright disputes that have followed other AI music generators.

For platforms and marketplaces built around AI-generated audio, watermarking and artist-style restrictions are becoming table stakes rather than a differentiator — the bigger competitive question is shifting toward how usable the output actually is in a real production pipeline, not just how it sounds in a demo. Structure control and image-driven prompting point in that direction: they’re aimed less at novelty generation and more at giving creators a repeatable way to steer a track toward a specific outcome.