Google DeepMind Releases Lyria 3.5 to Enhance Music Structure and Duration Control
Google DeepMind has launched the music generation model Lyria 3.5, focusing on improvements in music structure, lyric quality, instruction adherence, vocal performance, and song duration control, capable of directly generating complete songs of up to approximately 3 minutes.
The model continues to use a latent diffusion architecture, generating in the time audio latent space with training data annotated at different granularities. It has been trained through SFT and reinforcement learning with human and critic feedback, with all outputs embedded with SynthID watermarks; it shows significant improvements in audio clarity and lyric instruction adherence compared to its predecessor.
The model is initially integrated into Flow Music, supporting conversational creation, stem separation, remixing, and playlist generation, and can link with Veo to produce music videos, further integrating Lyria, Veo, and Gemini to form a complete music ecosystem from creation to distribution, concentrating funding and resources towards Google's proprietary creation platform.
Source: Public Information
ABAB AI Insight
Google DeepMind has previously entered the music generation space with the Lyria series, moving from early limited access to Lyria 3 being implemented in Gemini and Vertex AI, and then to Lyria 3 Pro supporting complete tracks of about 3 minutes. The path has always been to first validate the technology and then embed it into its own ecosystem, avoiding direct price competition with independent tools like Suno and Udio.
Capital and computing power continue to flow into the Flow Music and Gemini systems: through API openness and Spaces customization tools, Google is transforming model capabilities into chargeable creative workflows, while establishing traceable identifiers with SynthID to reduce copyright dispute risks, similar to its deployment logic of SynthID in the image and video domains.
In comparison, Stability AI and Meta's music models remain more in the research or open weights phase, while Google opts for closed iteration and platform binding, currently in a control period transitioning from short segments to complete songs and production tools, highly similar to Adobe Firefly's embedding strategy within creative software.
This essentially represents a restructuring of the industry chain: generative models are no longer just sold as API calls, but through duration control, stem export, and cross-modal linkage, consolidating the entry point, editing, and distribution into the same ecosystem, reducing reliance on external tools, and strengthening the platform's pricing power over the music production chain.
ABAB News · Cognitive Laws
The stronger the model, the more expensive the entry point
Watermarks are not protection, but control
Complete duration determines the ecological boundary