Lyria 2 – An AI music generation model launched by DeepMind

What is Lyria 2?

Lyria 2 is the third-generation AI music generation model developed by Google DeepMind, serving as a core component of the Vertex AI platform.
It is capable of producing professional-grade stereo audio at 48kHz, 24-bit quality, supporting a wide range of musical genres, including pop, rock, and classical, with the ability to blend styles seamlessly.

Lyria 2 expands upon the Lyria RealTime model, enabling real-time music creation and editing, allowing creators to trigger AI-driven compositions instantly through MIDI controllers or microphones.
It introduces three new features—“Create”, “Extend”, and “Edit”—for generating music from text prompts, extending existing audio clips, and adjusting the mood and style of music, respectively.

Key Features of Lyria 2

• High-Fidelity Audio Generation
  Lyria 2 generates professional-grade, 48kHz high-fidelity stereo audio, supporting various genres such as pop, classical, electronic, and more.

• Real-Time Creation and Adjustment
  Expanding on the Lyria RealTime model, it allows users to create and modify music in real-time, dynamically altering style, rhythm, and emotion—ideal for live performances and rapid prototyping.

• Text-Driven Composition
  With the new “Create” feature, users can generate music clips from simple text descriptions or lyrics, significantly lowering the barrier to music creation.

• Audio Continuation Generation
  The “Extend” feature enables the generation of extended content based on existing audio clips, assisting creators in expanding their musical pieces.

• Audio Editing and Style Fusion
  Through the “Edit” feature, users can modify the mood and style of audio clips via preset options or text prompts, and even merge different musical sections together.

• Digital Watermarking Technology
  All music generated by Lyria 2 is embedded with SynthID watermarking technology, ensuring the content’s AI origin is recognizable and helping to prevent copyright disputes.

Technical Principles

Multimodal Input and Parameterized Control

Lyria 2 supports various input formats, including text, MIDI, and audio clips.
Users can control music generation through descriptive prompts (e.g., “a cheerful jazz piano piece”) or parameters such as BPM and musical key.
The generation engine combines diffusion models and GANs (Generative Adversarial Networks) to enable style disentanglement and recombination.

Real-Time Interaction and Dynamic Creation

Lyria RealTime functionality leverages TensorRT acceleration and streaming inference, reducing generation latency to under 500ms.
This enables real-time music creation, allowing users to trigger compositions instantly via MIDI controllers or microphones.

Modular Creation Platform

Lyria 2 is integrated into the Music AI Sandbox, offering a modular toolchain from melody generation to chord progression.
It supports drag-and-drop functionality to combine different modules, significantly lowering the barrier for creative experimentation.

Project Resources

• Official Website: deepmind.google/technologies/lyria/

Application Scenarios

• Professional Music Production
  Generate high-fidelity, professional-grade stereo music across multiple genres for background music, sound effects, or full compositions.

• Creative Inspiration
  Quickly create music fragments from text descriptions or lyrics, helping artists break through creative blocks.

• Film and TV Scoring
  Generate dynamic scores that match visual moods based on script keywords, supporting real-time adjustments to boost production efficiency.

• Advertising Background Music
  Rapidly create customized music based on brand keywords and emotional tags, allowing ad teams to iterate faster and cut costs.

• AI Music Education
  Collaborate with institutions like Berklee College of Music and Stanford CCRMA to develop AI-driven music education courses, promoting innovation in music learning.