Curiosity · AI Model
Lyria 2
Lyria 2 is DeepMind's 2024-25 successor to Lyria, the research model behind Dream Track on YouTube Shorts. It generates up to several minutes of coherent music — melody, harmony, and instrumentation — from a text prompt, and ships as the engine behind Music AI Sandbox tools for professional musicians.
Model specs
- Vendor
- Family
- Lyria
- Released
- 2024-11
- Context window
- 4,096 tokens
- Modalities
- text, audio
Strengths
- Multi-minute coherent composition with structure and transitions
- Built-in SynthID watermarking for provenance
- Integrated into Google creator tools and partner pipelines
Limitations
- Closed-weights and largely product-gated — limited raw API access
- Vocal generation restricted by rights agreements
- Text-to-music benchmarks are inherently subjective
Use cases
- Soundtrack generation for short-form video
- Music AI Sandbox experiments for composers
- Royalty-tracked Dream Track covers in YouTube Shorts
- Prototyping film, game, and ad scores
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| Internal listener preference vs. Lyria 1 | ~65% preference | 2024-11 |
Frequently asked questions
What is Lyria 2?
Lyria 2 is Google DeepMind's second-generation text-to-music model. It generates multi-minute compositions from natural-language prompts and powers Music AI Sandbox and YouTube Shorts Dream Track features.
Can I use Lyria 2 via an API?
Lyria 2 is mostly product-gated inside Music AI Sandbox, Dream Track, and Vertex AI's generative-media APIs, rather than being directly available as an open API model.
How does Lyria 2 mark AI-generated audio?
Lyria 2 outputs are watermarked with Google's SynthID system, embedding an inaudible signal detectable by SynthID tooling for provenance.
Sources
- DeepMind — Lyria 2 and Music AI Sandbox — accessed 2026-04-20
- Google — SynthID overview — accessed 2026-04-20