Google DeepMind releases Lyria 3 for high-fidelity long-form music generation in Gemini

Lyria 3 marks a step-change in auditory AI: users can generate high-fidelity, long-form tracks up to three minutes from text or image prompts directly inside the Gemini app. Google says the model excels at understanding foundational musical elements — structure, key, rhythm — and at ensuring seamless transitions across a full composition rather than producing disjointed clips.

The safety mechanism is notable: SynthID watermarking is embedded so AI-generated content can be identified even after edits, an attempt to get ahead of provenance and copyright concerns that have dogged generative audio. That positions Google as taking a more cautious, traceable stance than some rivals in a legally fraught domain.

Competitively, Lyria 3 squares off against Suno, Udio, and Stability's audio efforts, with Google's advantage being native distribution inside Gemini's consumer surface and integration with its broader multimodal stack. Bringing three-minute, professional-grade output to a mainstream app lowers the barrier for creators considerably.

The open questions are familiar to the music-AI space: licensing of training data, how labels and artists respond, and whether watermarking survives real-world re-encoding and remixing. For now it's a clear signal that Google intends to compete across every generative modality — text, image, video (Gemini Omni Flash topped the Video Arena this week), and now audio.