Google Flow MusicにLyria 3.5を導入、音楽性・歌詞・ボーカル・クリエイティブ制御が向上We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control

Google DeepMind has released Lyria 3.5, the latest iteration of its generative music model, within Google Flow Music, marking another step in the company's push into AI-assisted music creation. The update matters because it targets the four areas that have most often limited AI-generated music: overall musicality, the coherence of generated lyrics, the naturalness of synthesized vocals, and the degree of hands-on control users have over the final output.

According to Google DeepMind's announcement, Lyria 3.5 brings meaningful improvements across all four dimensions. Better musicality suggests more convincing arrangements, rhythm, and instrumentation, while stronger lyric generation points to text that scans more reliably and stays on theme. Improved vocal quality addresses one of the persistent weak points in AI music, where synthetic singing can sound artificial or unstable across a track. The emphasis on creative control indicates that users are being given more direct influence over how a piece develops rather than simply accepting a single generated result.

Lyria 3.5 sits within Google Flow, the company's broader creative suite. Flow was introduced as a filmmaking-oriented environment that combined Google's models for video, image, and audio generation, with Veo handling video and Imagen handling stills. Adding a more capable music engine through Flow Music fits the platform's goal of letting creators assemble multiple media types in one place, so a generated score or song can be paired with generated visuals inside the same workflow.

The model builds on Google's earlier Lyria work. The original Lyria debuted through a collaboration with YouTube and was used in experiments such as Dream Track, which let selected creators generate short soundtracks in the style of participating artists. Later versions powered Google's Music AI Sandbox and MusicFX tools, aimed at musicians and hobbyists experimenting with text-to-music prompts. Lyria 3.5 appears to continue that trajectory, moving from short clips and stylistic experiments toward fuller, more controllable compositions.

The release lands in a competitive and increasingly scrutinized market. Independent startups such as Suno and Udio have drawn large user bases with text-to-song tools, and both have faced legal action from major record labels over the data used to train their systems. That context helps explain why established players like Google tend to stress provenance and safeguards. Google has said it applies SynthID, its watermarking technology, to AI-generated audio so that machine-made tracks can be identified, though the specific safeguards attached to Lyria 3.5 in Flow Music were not detailed in the summary.

Technically, models in this class are typically trained on large audio datasets and generate music either as waveforms or through intermediate token representations that are decoded into sound. Improvements in vocal quality and lyric coherence often reflect better alignment between the text a user supplies and the audio the model produces, as well as refinements in how the system handles longer time structures such as verses and choruses. Google has not published detailed benchmarks in the material available, so the scale of the gains relative to previous Lyria versions is best treated as the company's own characterization rather than an independently verified measure.

For creators, the practical takeaway is that Flow Music is being pitched as a more finished tool for producing songs with vocals, not just instrumental beds or background loops. The combination of stronger lyrics and vocals is likely aimed at users who want complete tracks, while the added creative control is meant to make results more repeatable and easier to steer.

Broader questions remain around licensing, artist compensation, and how platforms will label AI-generated songs as they circulate on streaming services. Those issues sit largely outside the scope of this particular update, but they form the backdrop against which every major AI music release is now judged. As with earlier launches, the real test for Lyria 3.5 will be how it performs in the hands of everyday users and whether the promised improvements hold up across a wide range of genres and prompts.