Space DJ: Navigating a Musical Universe (magenta.withgoogle.com)

Magenta has launched Space DJ, an interactive web app that turns music discovery into a spatial, real-time generative experience: you pilot a spaceship through a 3D “galaxy” where each star is a musical genre and nearby stars dynamically blend to produce continuous, on-the-fly music via the Lyria RealTime API. Clicking a star anchors genre weights, an auto-pilot mode creates evolving soundscapes, and a “High-Dimensional Neighbors” toggle reveals semantically similar genres that don’t appear close in the 3D projection. The result is an intuitive way to explore genre relationships and co-create sonic mashups with immediate auditory feedback. Under the hood, Space DJ maps text prompts for 300 selected genres (from a 1000-genre pool) into 768-d embeddings using MagentaRT’s MusicCoca embedder, reduces them to 128 dims with PCA, and projects them into 3D using UMAP (tunable in settings). The browser renders the constellation and controls with three.js, translating proximity into weighted text prompts (e.g., Deep House 0.7, Ambient Techno 0.3) which the Lyria RealTime model synthesizes with low latency. Development leveraged AI Studio and Cloud Run for iterative coding and secure deployment, proxying API requests to protect keys. For the AI/ML community this demonstrates a tight pipeline: embedding geometry + dimensionality reduction + interactive UI + real-time generative models—useful for research into embedding interpretability, low-latency synthesis, and creative interfaces.