Google DeepMind Rolls Out Lyria 3.5 in Flow Music
Google DeepMind has introduced Lyria 3.5 in Google Flow Music, saying the model improves musicality, lyrics and vocal quality while supporting variable song lengths. The update gives creators a newer model inside Google's song-generation workspace, but Google has not published comparative benchmarks for the claimed quality gains.
Google DeepMind has introduced Lyria 3.5 in Google Flow Music, updating its music-generation system with what the company describes as improvements to musicality, lyrics and vocal quality. DeepMind's product page also lists variable song lengths among the model's changes.
What changed in Lyria 3.5
The official Lyria page presents 3.5 as the latest model in the family and provides examples spanning several genres, languages and vocal styles. It says creators can use prompts to shape elements such as tempo, arrangement, instruments and vocal character. The broader Lyria workflow can turn text or an image into a track, and Google's page says generated pieces can run for up to three minutes.
Music Business Worldwide reported the rollout on July 30 and said the model is available inside Google Flow Music, the company's music-creation workspace. That positioning matters: this is a model update delivered through an existing creation product, not the announcement of a separate consumer music service.
Google's public materials do not provide controlled listening tests, preference-study results or head-to-head benchmarks against systems such as Suno. Claims about more realistic or nuanced vocals should therefore be read as Google's product characterization rather than independently measured performance.
Controls, provenance and limits
DeepMind says Lyria can generate lyrics and vocals in different languages, follow structural and stylistic instructions, and export professional-grade audio. The product page also says all generated tracks carry SynthID, Google's imperceptible watermark for identifying music made or edited with its AI.
The same page acknowledges that the system still has limitations and advises creators to check whether generated tracks fit their intended result. It also describes filtering and data-labeling work intended to reduce harmful lyrics, while stopping short of publishing new technical evaluation data specifically for Lyria 3.5.
For data and AI teams, the practical signal is the product integration: Google is iterating on generative audio quality and controllability inside a creator-facing workflow. The important questions for later evaluation are measurable vocal fidelity, prompt adherence, edit consistency and whether provenance signals survive common post-production steps.
Key Points
- 1Google DeepMind describes Lyria 3.5 as improving musicality, lyrics and vocal quality, with variable song lengths.
- 2The model is available through Google Flow Music, where creators can steer structure, style, instrumentation and vocals.
- 3Google says generated tracks include SynthID, but it has not published comparative benchmarks for the 3.5 quality claims.
Scoring Rationale
Lyria 3.5 is a meaningful update to Google's creator-facing generative-audio stack, with claimed improvements in vocals, lyrics and musical control. Practitioner value is moderated by the absence of comparative benchmarks or a new technical evaluation specific to version 3.5.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
