Google DeepMind Releases Lyra 3.5 Music Model Generating Full 3-Minute Songs

Lyra 3.5 utilizes latent diffusion to produce complete 3-minute tracks with improved lyric compliance and vocal quality, debuting within the Flow Music ecosystem.

Woofun AI reports that Google DeepMind has launched Lyria 3.5, a music generation model capable of producing full songs up to three minutes in duration. The system employs a latent diffusion architecture within the time-audio latent space, trained on audio with text annotations and refined via SFT combined with reinforcement learning using human and Critic feedback. All outputs include SynthID watermarks, though specific training data scales and sources remain undisclosed. Lyria 3.5 integrates into Flow Music, enabling conversational creation, stem separation, remixing, and collaboration with Veo for music videos, while supporting plugin development and custom DAWs to unify creation, editing, and distribution.