BlockBeats News, July 30, Google DeepMind released the new generation music generation model Lyria 3.5, focusing on enhancing music structure, lyrics quality, instruction adherence, vocal performance, and song duration control capability. It can directly generate complete songs up to 3 minutes long, rather than just audio snippets of a few seconds.

According to the introduction, Lyria 3.5 still adopts the latent diffusion architecture, generating diffusion in the temporal audio latent space. The training data consists of audio with text annotations of various granularities and undergoes post-training through SFT and reinforcement learning combined with human and Critic feedback, with all generated content embedded with SynthID watermark. However, Google did not disclose the scale, source, and quantitative benchmark of the training data, only stating that compared to Lyria 2, there has been a significant improvement in audio clarity and lyric instruction adherence.

In addition, Lyria 3.5 is first integrated into Flow Music, which supports conversational music creation, stem tracking, remixing, music publishing, playlist generation, and can be linked to Veo for music video generation. It also supports audio plugin development, music games, and custom DAWs, further integrating Lyria, Veo, and Gemini to build a complete ecosystem covering music creation, editing, publishing, and distribution.