Lyria 3.5 makes Google Flow Music more controllable, not just more musical

Google DeepMind has released Lyria 3.5, a new music-generation model now powering Google Flow Music. Google’s announcement is dated July 29, 2026, and Flow Music’s official launch post was published at 16:00 UTC. 1 2

The headline is not simply higher-fidelity audio. Lyria 3.5 is aimed at the parts of music generation that determine whether a result can be edited into a usable song: richer and more complex melodic structures, arrangements that move more naturally, stronger lyric quality, and better adherence to both simple and complex creative instructions. 1

The model’s vocal upgrade is the most visible one. Google says Lyria 3.5 produces more realistic, expressive, and emotionally nuanced singing, with improved pronunciation. Flow Music’s launch post adds BPM control and exportable stems for full-length songs, giving users more control over tempo and a way to carry separate musical parts into a downstream production workflow. 1 2

That makes the practical difference less about generating one impressive demo and more about reducing the cost of iteration. A producer can specify a tempo, shape a longer song, and work with stems instead of treating every generation as a finished, indivisible audio file. The Flow Music workspace is available to try, and its product page presents full-length song creation, remixing, stem splitting, and audio effects as part of the same environment. 3

Google’s model card reports human and automated evaluations covering music quality, vocal quality, audio fidelity, and prompt adherence, with comparisons against Lyria 2. It describes significant gains in audio fidelity and better prompt adherence when lyrics are included, but the public card does not provide numerical benchmark scores. That leaves the release with a clear direction of improvement, not a standardized score that can be cleanly compared with Suno, Udio, or other music models. 4

The card identifies text as the input and music audio plus lyrics as outputs. It says the model is distributed through Flow Music and can be made available to downstream providers through an API, but does not describe a general self-hosted release. Its stated boundaries are product and policy limits: Google prohibits uses involving illegal or dangerous activity, rights violations, harmful content, deception, or threats to service security. Google also lists safety filtering and SynthID watermarking among its mitigations. 4

The significance is therefore specific: Lyria 3.5 pushes Google’s music offering toward a controllable production tool. Whether its vocals and arrangements hold up outside curated examples will depend on hands-on testing, but BPM control and stems are concrete workflow features rather than another quality claim.

Read Google’s announcement · Read the model card