Google DeepMind has released the Lyria 3.5 music generation model, which supports generating full songs up to three minutes in length.

Google DeepMind has launched its next-generation music generation model, Lyria 3.5, with key enhancements to music structure, lyric quality, instruction adherence, vocal performance, and song duration control. The model can directly generate full songs up to 3 minutes long, rather than just tens of seconds of audio clips. Per the announcement, Lyria 3.5 still adopts the latent diffusion architecture, generating outputs via diffusion in the temporal audio latent space. Its training data comprises audio paired with text annotations of varying granularities, while post-training is conducted through supervised fine-tuning (SFT) and reinforcement learning that integrates both human and critic feedback. All generated content is embedded with SynthID watermarks. However, Google did not disclose the scale, source, or quantitative benchmarks of the training data, only noting that it has achieved significant improvements in audio clarity and lyric instruction adherence compared to Lyria 2. Furthermore, Lyria 3.5 is initially integrated into Flow Music, which supports conversational music creation, stem separation, remixing, music publishing, playlist generation, and can connect with Veo to produce music videos. It also enables the development of audio plugins, music games, and custom DAWs, further integrating Lyria, Veo, and Gemini to build a complete ecosystem spanning music creation, editing, publishing, and distribution.