Google launched Magenta RealTime — a new neural network for creating music in real-time. Instead of generating entire tracks, the model creates music 2 seconds at a time, which allows it to be used even on low-end hardware.
The model is trained on 190,000 hours of instrumental music from open sources (without vocals) and contains 800 million parameters. It analyzes the last 10 seconds of the generated melody, ensuring musical coherence. Generating one fragment takes about 1.25 seconds when running in Google Colab (free tier).