Google Lyria 3.5 music model adds section-level track edits

Google released the Lyria 3.5 music model with tighter editing controls for AI-generated tracks, including a way to revise individual sections rather than regenerate an entire song. The release matters for creative AI operators because it shifts Google’s music tool from one-shot generation toward an iterative production workflow, though the company has not disclosed training data details for the new model.

According to Google, Lyria 3.5 produces more natural melodies, stronger lyrics and more realistic vocals with clearer pronunciation than prior versions. The model can generate tracks between 30 seconds and three minutes, and users can set tempo and track length with more precision, the company said.

Lyria 3.5 is available through Google Flow Music, the company’s music-generation product. The most notable addition is a feature Google calls “Selective Section Painting,” which lets users alter a specific passage in a track or expand a short melody into a full song without discarding the rest of the work.

What can Lyria 3.5 do?

Lyria 3.5 is Google’s music-generation model for creating and revising generated audio tracks inside Flow Music. Google says users can now adjust parts of a composition, including vocals, drums, bass and other elements, with more control over tempo and duration.

That is a practical change for anyone trying to use generated music in a workflow rather than as a novelty output. If a chorus, drum section or vocal passage is wrong, section-level editing gives users a way to repair the piece while preserving the parts that already work. Google did not disclose usage numbers for Flow Music or say whether Lyria 3.5 changes pricing or access terms.

The company’s claims about sound quality are broad, and the announcement does not provide benchmark results or third-party evaluations. Better vocal realism and clearer pronunciation are measurable only in use, especially for creators who need consistent lyrics, timing and mix structure across multiple revisions.

What did Google say about Lyria training data?

Training data remains the unresolved issue around Google’s music models. When Google introduced Lyria 3, the company said that model had been trained on material YouTube and Google had the right to use under terms of service, partner agreements and applicable law. Google did not give a detailed list of the training materials.

Asked about training data for Lyria 3.5, Google did not immediately respond, The Decoder reported. That leaves the new model in the same broad category as many generative media systems: more capable tools are arriving before companies provide granular disclosures about the catalogs and rights structures behind them.

For Google, Lyria 3.5 extends its push into generative audio at a time when model providers are trying to make creative tools feel less disposable. For users, the relevant test is whether Flow Music can now support revision-heavy production work, not just generate a plausible first draft.

This story draws on original reporting from The Decoder.