Google launches Lyria 3.5 in Flow Music with better vocals and control

Google has launched Lyria 3.5 inside Google Flow Music, giving users access to its newest music-generation model from July 29. The update is focused less on adding another novelty and more on improving the parts that determine whether an AI-generated track feels usable: melody, lyrics, vocals, pronunciation, tempo and duration.

According to Google, Lyria 3.5 can generate richer, more complex melodic structures with a more natural sound. Lyrics should follow prompts more closely and show better structural awareness, while vocals are designed to sound more expressive and emotionally nuanced. Google also says pronunciation has improved. Users can now steer tempo and output length more directly, two controls that are essential when a song must fit a particular scene, edit or arrangement.

The model is rolling out through Flow Music, Google's generative music environment for creating, remixing and sharing songs. That positioning matters. Rather than presenting Lyria only as a model or a short audio demo, Google is placing it inside a workflow where people can develop ideas, revisit tracks and combine generation with other creative decisions.

For musicians, video makers and other creators, the practical value is speed. A written idea can become a structured musical sketch without recording a singer or assembling a full production team. Better lyrics and vocals may make early demos more convincing; tempo and duration controls can reduce the amount of editing needed afterward. Human judgment remains central for taste, arrangement, originality and whether a track suits its intended context.

The announcement is also a competitive signal. AI music products are moving beyond one-click song generation toward broader creative platforms. Quality alone is no longer enough: creators also need control, iteration, organisation and a clear way to turn an experiment into something usable. By connecting Lyria 3.5 directly to Flow Music, Google can learn from real creative workflows while keeping the model and interface closely integrated.

There are still open questions. Google's short release does not provide independent listening tests, benchmark results, detailed availability by region or new information about training data and rights. Claims about musicality and realism therefore need to be tested in practice. Even with that caveat, Lyria 3.5 shows where generative music is heading: not merely producing audio from a prompt, but giving creators more deliberate control over the shape and performance of a song.