Google release ‘Lyria 3 Pro’

Expanding on last month’s release of Lyria 3, Google DeepMind has officially launched Lyria 3 Pro, a professional-grade music generation model capable of composing full three-minute songs with advanced structural control.

While the standard Lyria 3 (released February 18) was optimized for 30-second social media clips and jingles, the “Pro” variant allows users to architect entire musical narratives by prompting for specific intros, verses, choruses, and bridges.

1. Key Features of Lyria 3 Pro

The new model moves beyond “snippet” generation to act as a collaborative musical architect.

• Extended Duration: Supports high-fidelity (48kHz stereo) tracks up to 180 seconds (3 minutes), a sixfold increase over the base model.
• Structural Awareness: Users can now use “component-based prompting,” such as: “Start with a solo piano intro, transition into an upbeat pop verse, and climax with a heavy synth chorus.”
• Multimodal Input: Like its predecessor, Lyria 3 Pro can generate music based on text prompts, images, or even short video clips, matching the “vibe” and atmosphere of the visual content.
• Agentic Iteration: Integrated into the new ProducerAI platform, the model allows artists to iterate on specific stems (vocals, drums, or melody) without regenerating the entire track.

2. Availability & Subscription Tiers

Google is rolling out Lyria 3 Pro across its entire ecosystem, with specific daily limits for paid Gemini subscribers.

Platform Rollout:

• Gemini App (Web/Mobile): Live as of March 25 for paid subscribers.
• Google Vids: Integrated this week for Workspace and Education customers.
• Vertex AI & Google AI Studio: Public preview available for enterprise and developer use.

3. Safety and Training: The “SynthID” Standard

To address the growing concern over “AI slop” and copyright, Google has implemented several strict guardrails:

1. SynthID Watermarking: All audio generated by Lyria 3 Pro contains an inaudible digital watermark. This remains detectable even if the file is compressed to MP3 or re-recorded through a microphone.
2. No Direct Mimicry: If a prompt includes a specific artist’s name (e.g., “Make a song like Taylor Swift”), the model is programmed to take “broad stylistic inspiration” without copying the artist’s unique vocal likeness or signature melodies.
3. Authorized Training: Google maintains that the model was trained exclusively on audio data that Google has the right to use through partner agreements or YouTube’s licensing terms.

4. Technical Comparison: Lyria 3 vs. Lyria 3 Pro

“Lyria 3 Pro is a genuine shift from ‘helper’ to ‘architect,'” noted one tech analyst. “It enables full compositions that could realistically serve as a film score or a radio-ready track with minimal human intervention.”