Google's Lyria 3.5 lets users edit song sections without restarting

Google has unveiled Lyria 3.5, the latest iteration of its music generation model, available now through Google Flow Music. The model delivers improvements across multiple dimensions: it generates more natural-sounding melodies, higher-quality lyrics, and more realistic vocals with clearer pronunciation. Users can control tempo and track length more precisely, with output ranging from 30 seconds to 3 minutes.

The headline feature is "Selective Section Painting," a new tool that fundamentally changes how users interact with the model. Instead of regenerating an entire track to modify a single section, users can now edit specific parts of a song in place. The feature also enables users to take short melodies and expand them into full songs without restarting the generation process. Beyond structural editing, users gain fine-grained control over individual sonic elements—they can adjust the tempo and duration of vocals, drums, bass, and other components independently.

When Google launched Lyria 3, the company stated that it had trained the model on materials covered by YouTube and Google's terms of service, partner agreements, and applicable law. However, Google declined to disclose detailed information about the training dataset. When asked specifically about the training data for Lyria 3.5, Google did not immediately respond, leaving the provenance of the new model's training materials unclear.

Google's release of Lyria 3.5 represents an incremental improvement over the previous Lyria 3 model, focusing on both output quality and user control. The addition of 'Selective Section Painting' addresses a real friction point in AI music generation—the need to restart from scratch when making changes to a song. By allowing users to edit specific sections and adjust individual element parameters (tempo, duration of vocals, drums, bass), Google is shifting the interaction model from "generate and accept" to "generate, refine, and iterate."

The training data question remains unresolved. Although Google disclosed that Lyria 3 was trained on materials covered by YouTube and Google's terms of service, partner agreements, and applicable law, the company has declined to provide specifics about Lyria 3.5's training sources. This lack of transparency mirrors broader questions in the AI industry about data sourcing and rights clearance for generative music models.

No comments yet. Be the first to share your thoughts!