Google DeepMind releases Lyria 3.5 music generation model, which supports generating full songs of up to 3 minutes

Comparatively, Google DeepMind released Lyria 3.5, a next-generation music generation model, which focuses on improving the music structure, quality of lyrics, command compliance, vocal performance, and the ability to control song length. It can directly generate a complete song of up to 3 minutes, rather than just tens of seconds of audio.

According to reports, Lyria 3.5 still uses a latent diffusion architecture to diffuse and generate in the temporal audio latent space. The training data is audio with text annotations of different granularity, and post-training is performed through SFT and reinforcement learning combined with human and critical feedback. SynthID watermarks are embedded in all generated content. However, Google did not disclose the scale, source, and quantified Benchmark of the training data, which only indicated a significant improvement in audio clarity and compliance with lyric instructions compared to Lyria 2.

In addition, Lyria 3.5 was first integrated into Flow Music, which supports interactive music creation, stem splitting, remix, music publishing, playlist generation, and can also link Veo to generate music videos. It also supports the development of audio plug-ins, music games, and custom DAWs to further integrate Lyria, Veo, and Gemini to build a complete ecosystem covering music creation, editing, publishing, and distribution.