We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·May 20·5 min read
Stable Audio 3
TL;DR
Audio generation models struggled to produce high-quality, controllable music and sound effects at scale. Stability AI released Stable Audio 3, a generative model that creates audio from text descriptions and musical scores, using diffusion (a technique that iteratively refines noise into coherent output).
✦ Why It Matters
Engineers can now integrate higher-quality text-to-audio generation into applications without training custom models from scratch.
Key Takeaways
- Utilize Stable Audio 3 for efficient audio generation and editing.
- Leverage inpainting features for precise audio modifications.
- Run models on consumer-grade hardware for quick audio production.
- Access released weights and training pipelines for practical implementation.
How It Works
Stable Audio 3 employs latent diffusion models that generate audio by projecting it into a compact latent space using a semantic-acoustic autoencoder. This approach allows for efficient diffusion-based generation while ensuring high audio fidelity and maintaining semantic coherence in the output.